Open Source / 開源
-
Tools ENOne Message, Root Shell: Unpatched CVSS 9.8 RCE in LMCache Exposes vLLM Inference Stacks
CVE-2026-105192 turns LMCache's multiprocess ZeroMQ port into an unauthenticated pickle-deserialization RCE that runs as root in official containers — and no fixed version exists.
-
Tools 中一則訊息、一個 root 權限：LMCache 未修補的 CVSS 9.8 遠端程式碼執行漏洞暴露 vLLM 推論叢集
CVE-2026-105192 讓 LMCache 多程序模式的 ZeroMQ 埠號變成未經身分驗證的 pickle 反序列化 RCE，在官方容器中以 root 執行——而且目前沒有任何修補版本。
-
Tools ENFive AI Targets, Five Falls: Pwn2Own Ireland Puts the AI Stack on the Hacking Stage
At Pwn2Own Ireland in Cork, researchers collected $388,500 and 32 zero-days on day one alone — and every AI Infrastructure target from OpenAI Codex to LiteLLM, Chroma, Dynamo and Oracle's Autonomous AI Database fell to attack.
-
Tools 中五個 AI 目標全數淪陷：Pwn2Own 愛爾蘭把 AI 基礎設施搬上駭客舞台
在愛爾蘭科克舉行的 Pwn2Own 競賽中，研究人員僅第一天就領走 38.85 萬美元賞金與 32 個零日漏洞——從 OpenAI Codex、LiteLLM 到 Chroma、Dynamo 與 Oracle Autonomous AI Database，AI 基礎設施類別的每個目標都被攻陷。
-
Tools EN75GB of RAM Later: Microsoft Puts a 137B-Parameter Coding Agent on Your Laptop
At its Windows and Surface event, Microsoft announced that GitHub Copilot will soon route tasks between cloud and a local, quantized MAI Code 1.1 Flash — a 137B-parameter MoE coding model that fits in 53GB — sandboxed by the new open-source Microsoft Execution Containers.
-
Tools 中75GB 記憶體之後：微軟把 137B 參數的程式代理放上你的筆電
在 Windows 與 Surface 發表會上，微軟宣布 GitHub Copilot 將自動在雲端與本機之間調度任務——本機端是量化到 53GB、可塞進筆電的 137B 參數 MoE 程式模型 MAI Code 1.1 Flash，並以新的開源沙箱庫 Microsoft Execution Containers 加以隔離。
-
Industry ENSixteen in One Month: Nikkei Data Shows China's Labs Now Set the Global Release Clock
Chinese developers shipped 16 models in September as the average US–China release cycle compressed from 125 to 44 days — and Beijing shows no sign of slowing down while Anthropic's CEO calls for pacing.
-
Industry 中一個月十六款模型：日經數據揭示中國實驗室正主宰全球 AI 的發布節奏
中國開發商九月一口氣發布 16 款模型，美中平均發布週期從 125 天壓縮至 44 天——正當 Anthropic 執行長呼籲放慢腳步之際，北京沒有展現任何減速跡象。
-
Models ENEight Milliseconds, Zero Tokens: Liquid AI Open-Sources Its d1 Decision Models for the Edge
Liquid AI releases open-weight d1-3B and multimodal d1-omni-600M — models that answer in one forward pass with no output tokens, running on everything from an RTX 4090 to a Jetson Orin Nano.
-
Industry ENFrom 24 Million Clones to the Boardroom: Nous Research Confirms $90M Series B, Launches Hermes for Businesses
The open-source Hermes Agent maker has closed a $90 million Series B at a $1.5 billion valuation and is heading to the enterprise with private, self-hosted AI agents.
-
Industry 中從 2,400 萬次克隆走向企業市場：Nous Research 確認 9,000 萬美元 B 輪融資，推出 Hermes for Businesses
開源 AI 代理 Hermes 的開發商以 15 億美元估值完成 9,000 萬美元 B 輪融資，並推出強調資料隱私與自主部署的企業版 AI 代理產品。
-
Models ENNinety Percent Off: Claude Haiku 5.5 Completes the 5.5 Family and Resets the Small-Model Price Floor
Anthropic ships Claude Haiku 5.5 at up to 90% below Haiku 4.5 pricing, with first-ever effort controls for the Haiku class, huge agentic benchmark jumps, and a system card that openly discloses safety regressions.
-
Models 中降價九成：Claude Haiku 5.5 補齊 5.5 家族，重寫小模型價格底線
Anthropic 於 10 月 7 日推出 Claude Haiku 5.5，短提示價格最低較 Haiku 4.5 便宜 90%，首度加入 effort 控制，代理式基準測試大幅躍進，系統卡更坦承揭露多項安全退步。
-
Tools ENThe Chip That Wrote Itself: openTPU Runs Qwen3 on an FPGA Its Agents Designed
A single GitHub committer claims AI agents designed a full-stack inference accelerator — RTL, ISA, compiler and all — and it now runs ten open models bit-exactly on a $200 Kintex-7 FPGA card.
-
Tools 中自己設計自己的晶片：openTPU 讓 AI Agent 打造的 FPGA 加速器跑起 Qwen3
一位 GitHub 開發者聲稱，AI agents 設計了整套推論加速器——從 RTL、指令集到編譯器——如今它在一張約 200 美元的 Kintex-7 FPGA 卡上，位元級精確地跑著十個開源模型。
-
Research ENThe Defender That Fights Back: AdvSim2Real Co-Evolves Web Agents With Their Attackers
MBZUAI, Amazon, and MIT researchers co-evolve a task curriculum, an injection adversary, and a web agent inside a frozen world model — lifting real-browser success from 25.6% to 44.4% while cutting prompt-injection losses.
-
Research 中會反擊的防禦者：AdvSim2Real 讓網頁 Agent 與攻擊者共同演化
MBZUAI、Amazon 與 MIT 的研究團隊在凍結的世界模型中，讓任務課程、注入攻擊者與網頁 Agent 三方共同演化——真實瀏覽器的成功率從 25.6% 提升到 44.4%，同時大幅降低提示注入的傷害。
-
Models ENOne Model, Every Medium: Google's EmbeddingGemma 2 Puts Multimodal Search in Your Pocket
Google DeepMind's EmbeddingGemma 2 is a 740M-parameter open model that maps text, code, images, video, and audio into one 768-dimensional space — in 191MB of RAM.
-
Models 中一個模型搞定所有媒體：Google EmbeddingGemma 2 把多模態搜尋裝進你的口袋
Google DeepMind 的 EmbeddingGemma 2 是一個 7.4 億參數的開源模型，將文字、程式碼、圖片、影片與音訊映射到同一個 768 維向量空間，文字模式僅需 191MB 記憶體。
-
Research EN722 Papers, One Model, Zero Human Authors: OpenAI's Largest Math Release Stress-Tests Verification Itself
OpenAI has published 722 machine-generated mathematics manuscripts across 372 research families — its largest AI math release yet, with Lean proofs for many results and a verification bottleneck the field never planned for.
-
Research 中722 篇論文、一個模型、零位人類作者：OpenAI 史上最大數學發布，考驗的是「驗證」本身
OpenAI 公開發布 722 篇由內部未發表前沿模型生成的數學研究手稿，分屬 372 個研究系列——多數成果附帶 Lean 形式化證明，卻也讓整個數學界面臨前所未有的驗證瓶頸。
-
Industry ENOne Login for Every Store: Meta and Sierra Publish the Personal Agent Protocol
Meta and Sierra's new open standard gives personal AI agents an OAuth-based way to authenticate with businesses — Walmart, Shopify, Stripe and Genesys are in, Visa's rival protocol is not going away.
-
Industry 中一組帳號走遍所有商店：Meta 與 Sierra 發布 Personal Agent Protocol
Meta 與 Sierra 发布的開放標準，為個人 AI 代理提供以 OAuth 為基礎的商家身分驗證機制 — Walmart、Shopify、Stripe、Genesys 已加入，而 Visa 的競爭協議也沒有退場跡象。
-
Policy ENThree Tiers, One Program: Anthropic Opens Mythos-Class Cyber Models to Vetted Defenders
Anthropic has rebuilt its Cyber Verification Program into three access tiers — Defense, Red Team, and Specialized — folding in Project Glasswing and opening Claude Mythos 5.1 to security teams, alongside the first hard numbers: 129,000+ verified vulnerabilities found since April.
-
Policy 中三層級、一套制度：Anthropic 將 Mythos 級網安模型開放給通過審查的防禦者
Anthropic 將「網路安全驗證計畫」（CVP）擴編為 Defense、Red Team、Specialized 三個存取層級，Project Glasswing 併入其中，Claude Mythos 5.1 首度有了常態申請管道；同時公布首批成果：4 月以來已驗證超過 12.9 萬個漏洞。
-
Industry ENThirty to One: The Lopsided Flow That Explains the US-China AI Talent Race
A Carnegie Endowment study reveals that for every 30 Chinese AI researchers in the US, only one made the reverse journey to China — Beijing's open-door visa push is landing, but almost nobody is walking through it.
-
Industry 中三十比一：一個數字看懂美中 AI 人才爭奪戰的真相
卡內基基金會研究顯示，每 30 位在美國工作的中國 AI 研究者，只有 1 人反向前往中國——北京大開國門搶人才，卻幾乎沒有人走進來。
-
Models ENOne Trillion Parameters, 49B Active: Mistral's Le Chonk Is Europe's Boldest Open-Weight Bet Yet
Mistral Large 4 'Le Chonk' — a 1T-parameter, 49B-active multimodal MoE trained on 3,800 Grace Blackwell GPUs in Europe — enters public preview with weights due October 27, posting 82% on vulnerability reproduction and second place in blind coding evals behind only Claude Opus 5.
-
Models 中一兆參數、490 億活躍參數：Mistral 的 Le Chonk 是歐洲迄今最大膽的開源權重豪賭
Mistral Large 4「Le Chonk」——一個在歐洲以 3,800 顆 Grace Blackwell GPU 從零訓練、1 兆參數、490 億活躍參數的多模態 MoE 模型——進入公開預覽，權重將於 10 月 27 日釋出；在漏洞重現測試拿下 82% 全場最高，盲測編碼評比僅次於 Claude Opus 5。
-
Industry ENOversubscribed: DeepSeek's Pre-IPO Round Balloons Past $12 Billion as Tencent and CATL Pile In
Bloomberg reports DeepSeek is closing in on at least 80 billion yuan (US$12B) — nearly double its own target — with signed term sheets pointing to a final tally near 100 billion yuan ahead of an early-2027 IPO.
-
Industry 中超額認購：DeepSeek 上市前融資暴增至 120 億美元，騰訊、寧德時代領銜進場
彭博報導 DeepSeek 最新一輪融資接近拿下至少 800 億人民幣（約 120 億美元），幾乎是原訂目標的兩倍；已簽署的投資條款顯示最終金額可能逼近 1,000 億人民幣，為 2027 年初 IPO 鋪路。
-
Industry ENMoonshot Closes the Book on Private Markets: Final Round Lands at $50B, Hong Kong IPO of Up to $5B Eyes Q1 2027
The Kimi K3 maker has wrapped its last private round at a $50 billion valuation and is sizing a Hong Kong listing of up to $5 billion for early 2027 — the biggest China AI flotation since the bubble found its footing.
-
Industry 中月之暗面告別一級市場：最後一輪私募以 500 億美元收官，香港 IPO 最高募 50 億美元劍指 2027 年第一季
Kimi K3 的開發商月之暗面已完成估值 500 億美元的最後一輪私募融資，並規劃於 2027 年初在香港上市、最高募資 50 億美元——這將是中國 AI 公司迄今最大規模的公開發行之一。
-
Policy ENInvisible Ink, Statistical Signature: OpenAI's textGrain Watermark Comes to EU ChatGPT Text
OpenAI's textGrain embeds an invisible statistical watermark in EU ChatGPT and Codex output to comply with the EU AI Act — strong on long prose, fragile under editing, and unlike Anthropic, optional for API users worldwide.
-
Policy 中看不見的墨水，統計的簽名：OpenAI 的 textGrain 浮水印登陸歐盟 ChatGPT 文字
OpenAI 的 textGrain 將在歐盟的 ChatGPT 與 Codex 輸出中嵌入隱形統計浮水印以符合歐盟 AI 法案——長文偵測強、編輯後脆弱，且與 Anthropic 不同，API 用戶全球皆為選用制。
-
Models ENBeam Lands: Reflection AI Ships a 501B-Parameter Open-Weight Model That Matches GLM-5.2 on a Fraction of the Compute
Reflection AI has officially unveiled Beam, a 501B-parameter sparse MoE model with 23B active per token, SWE-bench Verified at 80.9, and reasoning on par with Z.ai's GLM-5.2 using 3-4x less inference compute — the first credible American answer to China's open-weight dominance.
-
Models 中Beam 正式登場：Reflection AI 發布 501B 參數開放權重模型，以更少算力追平 GLM-5.2
Reflection AI 正式發表 Beam——總參數 501B、每 token 僅啟動 23B 的稀疏 MoE 模型，SWE-bench Verified 達 80.9，推理表現與 Z.ai 的 GLM-5.2 相當但推論算力只需 3-4 分之一，是美國陣營對中國開源模型壟斷局面的首次強力回應。
-
Models ENThree Percent: Bloomberg Intelligence Says DeepSeek Has Closed the US–China AI Gap to a Record Low
DeepSeek's V4.1 Flash scored 81.1 on LiveBench — sixth globally, a record for a Chinese model — cutting the US performance lead to ~3% from 15% in January, and raising hard questions about export controls and the profitability of China's 1,100-model price war.
-
Models 中三％的距離：Bloomberg Intelligence 指出 DeepSeek 已把美中 AI 差距縮到史上最小
DeepSeek V4.1 Flash 在 LiveBench 拿下 81.1 分、全球第六，創中國模型新高，把美國領先幅度從年初的 15% 壓縮到約 3%——出口管制的戰略效果與中國上千個模型混戰的獲利難題，同時浮上檯面。
-
Models ENAmerica's Open-Weight Answer: Nvidia-Backed Reflection AI Prepares Its First Model to Challenge DeepSeek and Qwen
Axios reports that Reflection AI — the $25B startup founded by ex-DeepMind researchers — will ship its first open-weight model this month, aiming to pair American-made weights with Nvidia's 'AI factory' vision and break China's grip on open-source AI.
-
Models 中美國的開放權重反擊：Nvidia 投資的 Reflection AI 準備以首款模型迎戰 DeepSeek 與 Qwen
Axios 獨家報導：由前 DeepMind 研究員創立、估值 250 億美元的 Reflection AI，本月將推出首款開放權重模型，結合 Nvidia 的「AI 工廠」願景，挑戰中國在開源 AI 的主導地位。
-
Meta ENSpinning Rust Strikes Back: Toshiba Will Double HDD Capacity for AI Data Centers by 2027
Toshiba is spending ¥60B ($380M) to double hard-drive output for AI data centers by FY2027, targeting 30% capacity share — and the news instantly wiped 10-12% off Seagate and Western Digital.
-
Meta 中旋轉磁碟的逆襲：東芝將在 2027 年前把 AI 資料中心硬碟產能翻倍
東芝投入 600 億日圓（約 3.8 億美元），要在 2027 財年前將 AI 資料中心用硬碟產能翻倍、搶下 30% 容量市占——消息一出，Seagate 與 Western Digital 股價應聲重挫 10%。
-
Tools ENGems Are Dead, Skills Are Live: Google Starts the Gemini Skills Rollout in Workspace Today
Starting October 5, Google begins replacing Gemini Gems with stackable, Markdown-native Skills across Workspace — here is the full timeline, what changes for admins and users, and why the SKILL.md standard matters.
-
Tools ENGems 走向終點、Skills 正式上線：Google 今日起在 Workspace 推出 Gemini Skills
Google 自 10 月 5 日起以可堆疊、Markdown 原生的 Skills 取代 Gemini Gems，逐步覆蓋 Workspace——完整時程、管理員與使用者需要知道的事，以及 SKILL.md 開放標準為何重要。
-
Research ENSelf-Improvement for $150: MIT and Sakana AI's SIFT Cuts the Cost of Recursive Agents
An LLM-judge-guided tree search lets a coding agent rewrite itself to 35.1% on Polyglot in five hours on $150 of API credits — a tenth of the compute of prior methods.
-
Research 中150 美元的自我進化：MIT 與 Sakana AI 的 SIFT 把遞迴自我改寫 Agent 的成本砍到十分之一
用 LLM 評審引導的樹搜尋，讓 coding agent 改寫自身後在 Polyglot 拿下 35.1%，只花 5 小時與 150 美元 API 費用——僅為過往方法十分之一的算力。
-
Models EN16,379 Probes Later: Independent Audit Strips Jev of Its Frontier Badge
A black-box audit of TypeSafe AI's decision model Jev ran 16,379 live benchmark requests plus 3,331 follow-up probes and concluded it is not frontier at all — a small 4–9B active-parameter scorer whose 97.9% ARC-Challenge score comes from a race that finished years ago.
-
Models 中16,379 次探測之後：獨立審計撕下 Jev 的「前沿」標籤
一份針對 TypeSafe AI 決策模型 Jev 的黑箱審計，跑完 16,379 次基準測試請求與 3,331 次後續探測，結論是它根本不是前沿模型——而是一個約 40 至 90 億活躍參數的小型評分器，97.9% 的 ARC-Challenge 分數來自一場早已結束的競賽。
-
Tools ENA 125B Model on a 12GB Gaming Card: How Strata Splits 24,576 Experts Across Your Whole PC
The open-source Strata engine runs Qwen3.8-Flash-Next, a 125B-parameter MoE model, on a single 12GB consumer GPU with 64GB of RAM at 60-95 tokens per second. Here is how expert offloading and speculative drafting make it work.
-
Tools 中125B 模型塞進 12GB 遊戲顯卡：Strata 如何把 24,576 個專家拆到整台 PC 上
開源推論引擎 Strata 讓 125B 參數的 MoE 模型 Qwen3.8-Flash-Next 在一張 12GB 消費級顯卡加上 64GB 記憶體的普通 PC 上，以每秒 60-95 token 的速度運行。關鍵在於專家卸載與投機解碼。
-
Models ENBanned Twice, Shipped Anyway: PewDiePie's Ajax Puts a 9B Uncensored Agent on Home PCs
Felix 'PewDiePie' Kjellberg fine-tuned Alibaba's Qwen3.5-9B into Ajax, an always-on local agent for his Odysseus workspace — after OpenAI suspended his account twice over distillation.
-
Models 中被禁兩次照樣出貨：PewDiePie 的 Ajax 把 9B「無審查」代理裝進家用電腦
Felix「PewDiePie」Kjellberg 以阿里巴巴 Qwen3.5-9B 微調出 Ajax，作為 Odysseus 自架工作區的常駐本地代理——而在開發期間，他的 OpenAI 帳號因「蒸餾」兩度遭停權。
-
Meta ENThe Bounty That Ate Itself: Google Suspends OSS VRP Product Reports as AI Slop Buries Maintainers
On October 1, 2026, Google stopped accepting product vulnerability submissions to its Open Source Software Vulnerability Reward Program, blaming an avalanche of invalid AI-generated reports — the first mainstream bounty to publicly crack under machine-made bug spam, with a restart promised by Q1 2027.
-
Meta 中被 AI 廢掉的安全通報管道：Google 暫停 OSS VRP 產品漏洞回報，維護者被「幻覺報告」淹沒
2026 年 10 月 1 日，Google 宣布不再受理開源軟體漏洞獎勵計畫（OSS VRP）的產品漏洞回報，理由是大量無效的 AI 生成報告湧入——這是第一個公開被機器垃圾報告壓垮的主流賞金計畫，重啟時間承諾訂在 2027 年第一季。
-
Policy ENThe Appeal Court Blinked First: Eighth Circuit Pauses Minnesota's AI Nudification Ban as xAI's First Amendment Fight Escalates
The Eighth Circuit granted xAI an injunction pausing Minnesota's first-in-the-nation AI nudification law while its constitutional challenge proceeds — reversing a lower court and freezing the tool-targeting statute that could have shaped a national template.
-
Policy EN上訴法院先眨了眼：第八巡迴法院暫停明尼蘇達 AI 裸體化禁令，xAI 的言論自由之戰升級
第八巡迴上訴法院批准 xAI 的禁制令聲請，在憲法訴訟期間暫停明尼蘇達全美首部 AI 裸體化禁令——推翻下級法院的裁決，凍結這部可能成為全國範本的「管制工具本身」法規。
-
Meta ENFound by AI, Weaponized in a Day: Mythos-Discovered Rejetto HFS Flaw Under Active Attack
CVE-2026-61500, a Rejetto HFS session-forgery flaw discovered by Anthropic's Mythos model, is being exploited in the wild by a China-based actor within 24 hours of disclosure.
-
Meta 中AI 找到的漏洞，一天內就被武器化：Mythos 發現的 Rejetto HFS 缺陷已遭實際攻擊
CVE-2026-61500 是 Anthropic Mythos 模型發現的 Rejetto HFS 會話偽造漏洞，技術揭露後不到 24 小時就遭中國背景攻擊者實際利用。
-
Research ENSix Papers, Five Open Problems: Meta's Muse Spark Joins the Mathematicians
Meta AI Research published six mathematics papers co-authored with Muse Spark 1.1 and 1.2 in Thinking Mode — five answer previously open problems, from a sharp ellipsoid-fitting threshold to a 384-element counterexample in group theory, all produced through the ordinary meta.ai chat box.
-
Research 中六篇論文、五個未解難題：Meta 的 Muse Spark 正式加入數學家行列
Meta AI Research 發布六篇由數學家與 Muse Spark 1.1／1.2（Thinking Mode）共同完成的論文，其中五篇解答了先前懸而未決的開放問題——從高維橢球擬合的嚴格閾值，到群論中 384 階反例，全部透過一般的 meta.ai 聊天視窗完成。
-
Models EN78B Parameters, 3B Active, Apache 2.0: Aleph Alpha's Kolibri Lands as Germany's Sovereign Open-Weight Bet
On the Day of German Reunification, Aleph Alpha open-sourced Kolibri: a 78B-total/3.5B-active MoE trained on 20T tokens with 21.3% organic German data, 1M-token context, and abstention training for regulated industries.
-
Models 中78B 參數、3B 啟動、Apache 2.0 開源：Aleph Alpha 的 Kolibri 打響德國主權 AI 模型之戰
在德國統一紀念日這天，Aleph Alpha 開源了 Kolibri：一個 78B 總參數、3.5B 啟動參數的 MoE 模型，以 20T token 訓練、21.3% 原生德文資料、百萬級上下文，並內建拒答訓練，瞄準受監管產業。
-
Industry ENA Database for Every Agent: Supabase Raises $150M and Buys Turso
GIC leads a $150M round into Supabase just four months after its $500M Series F, and the open-source Postgres leader acquires Turso to serve the 70% of new databases now created by AI agents.
-
Industry 中每個 Agent 一個資料庫：Supabase 募資 1.5 億美元並收購 Turso
GIC 領投 Supabase 1.5 億美元新輪資金，距其 5 億美元 F 輪僅四個月；這家開源 Postgres 平台同時收購 Turso，因為平台上 70% 的新資料庫已由 AI Agent 建立。
-
Models ENA 560B Model for Free: Ant Group Puts Ling-3.1-flash on OpenCode
Ant Group's InclusionAI made its 560B-parameter MoE Ling-3.1-flash free on OpenCode, ranking No. 2 among open models on Mobile App Arena — with an unresolved context-window gap and an unconfirmed license.
-
Models 中560B 模型免費開放：螞蟻集團把 Ling-3.1-flash 送上 OpenCode
螞蟻集團 InclusionAI 將 560B 參數的 MoE 模型 Ling-3.1-flash 在 OpenCode 上免費開放，於 Mobile App Arena 開源模型中排名第二——但上下文視窗規格仍有落差、授權條款也尚未確認。
-
Tools EN2.7x Faster MoE Training, Fully Open: Ai2 Ships Olmo-core 3 Into the Trillion-Parameter Era
The Allen Institute's redesigned open training stack keeps experts GPU-resident, hits 858 TFLOP/s per B300, and has been benchmarked past 2.38 trillion parameters.
-
Tools 中2.7 倍速的 MoE 開源訓練堆疊：Ai2 推出 Olmo-core 3 進軍兆級參數時代
艾倫人工智慧研究院全新設計的開源訓練框架讓專家常駐 GPU、單卡吞吐達 858 TFLOP/s，並已完成 2.38 兆參數的壓力測試。
-
Tools ENSix Repos Against CUDA: DeepSeek and Huawei Open-Source the Ascend Software Stack
DeepSeek has open-sourced its Ascend software stack — TileLang, DeepGEMM, DeepEP, FlashMLA, TileKernels, and DeepSelect — giving Huawei's chips a CUDA-alternative toolkit that already powers most operators in DeepSeek V4 training.
-
Tools 中六個開源儲存庫對決 CUDA：DeepSeek 與華為開源 Ascend 軟體堆疊
DeepSeek 開源了完整 Ascend 軟體堆疊——TileLang、DeepGEMM、DeepEP、FlashMLA、TileKernels 與 DeepSelect——為華為晶片提供 CUDA 之外的完整工具鏈，且已在 DeepSeek V4 訓練中承擔多數算子。
-
Tools ENHalf the Memory, Two-Thirds the Price: DGX Spark 64GB Brings Local Agents Down to $4,999
NVIDIA's new 64GB DGX Spark cuts the entry price of a Grace Blackwell desktop to $4,999, runs 100B-parameter models on device, and clusters two units to 128GB via the new Sync Cluster Assistant.
-
Tools 中記憶體減半、價格省三分之二：DGX Spark 64GB 把本地 AI 代理的入門價降到 4,999 美元
NVIDIA 推出 64GB 版 DGX Spark，把 Grace Blackwell 桌上型 AI 主機的入門價砍到 4,999 美元，單機可跑 100B 參數模型，兩台透過新版 Sync Cluster Assistant 叢集可擴充到 128GB。
-
Tools ENYour Editor, Rewritten at Runtime: Anthropic Opens Claude Code Itself With TypeScript Mods
Anthropic's new mods turn Claude Code inside out — small TypeScript functions can now rewrite prompts, block tool calls, redact secrets, and redraw the UI, with built-in features like /diff already converted to mods.
-
Tools 中把編輯器放進你的手裡：Anthropic 用 TypeScript Mods 打開 Claude Code 本體
Anthropic 推出的 mods 讓 Claude Code 徹底開放——小型 TypeScript 函式可在執行時改寫提示、攔截工具呼叫、遮蔽機密、重繪介面，連內建的 /diff 都已改寫為 mod。
-
Models ENThe Face That Fooled Half the Room: Tavus Griffin Passes the Video Turing Test
Tavus's Griffin-Lite convinced 48% of study participants it was human on live one-minute video calls, and tops NVIDIA's VideoFDB leaderboard — but the details cut both ways.
-
Models 中騙過半數人類的臉：Tavus Griffin 通過視訊版圖靈測試
Tavus 的 Griffin-Lite 在一分鐘視訊通話中讓 48% 的受試者相信它是真人，並登上 NVIDIA VideoFDB 排行榜冠軍——但細節值得仔細檢視。
-
Tools ENStop Renting Intelligence: NinjaTech's SuperNinja Enterprise Bundles GPUs, Inference, and AI Employees Into One Fixed Bill
NinjaTech AI's SuperNinja Enterprise deploys an unmetered AI workforce on open-weight models inside the customer's own cloud — GPUs, inference, and software in one contract, at roughly one-tenth the cost of frontier-lab deployments.
-
Tools 中別再按 token 租智慧：NinjaTech 推出 SuperNinja Enterprise，把 GPU、推論基礎設施與 AI 員工打包成一張固定帳單
NinjaTech AI 的 SuperNinja Enterprise 把無限量計費的 AI 員工部隊部署在客戶自己的雲端、跑開放權重模型——GPU、推論與軟體一紙合約搞定，端到端成本約為前沿實驗室部署的十分之一。
-
Models ENThe Model That Can't Write: AWS Open-Sources Strands Decider 2B, a 2B-Parameter Decision Engine for Agents
AWS's Strands Labs took a Qwen3.5-2B torso, deleted the LM head, and shipped a decision model that answers in tens of milliseconds on a laptop — fully open, weights, data, and training scripts included.
-
Models 中不會寫字的模型：AWS 開源 Strands Decider 2B，專為 Agent 而生的 20 億參數決策引擎
AWS 的 Strands Labs 拿 Qwen3.5-2B 當骨幹、直接刪掉語言模型頭，做出一個在筆電上幾十毫秒就能回答的決策模型——權重、訓練資料與腳本全部開源。
-
Tools ENNo New Stack Required: MongoDB Turns Itself Into an AI Agent Runtime
At its NYC Investor Day, MongoDB launched Atlas Agent Engine, a unified execution, memory, and governance layer for production AI agents, alongside MongoDB 9.0 and the Atlas Infinite scale-out tier — betting the database itself becomes the agent platform.
-
Tools 中不需新堆疊：MongoDB 把自己變成 AI Agent 執行環境
MongoDB 在紐約 Investor Day 發表 Atlas Agent Engine——為生產環境 AI agent 提供統一的執行、記憶與治理層，並同步推出 MongoDB 9.0 與 Atlas Infinite 彈性擴展層，押注資料庫本身就是 agent 平台。
-
Tools ENOne Loop to Ship Them All: CoreWeave Forge Turns Production Runs Into Better Models
CoreWeave's new Forge development layer unifies run, observe, curate, improve and evaluate into one open environment, with a coding agent named ARIA and serverless RL that trains 1.4x faster at 40% lower cost.
-
Tools 中一個迴圈走天下：CoreWeave Forge 讓生產環境的運行直接餵養下一代模型
CoreWeave 全新開發層 Forge 把運行、觀測、篩選、改進與評估整合進單一開放環境，搭配編碼代理 ARIA，以及速度快 1.4 倍、成本降 40% 的無伺服器 RL 訓練。
-
Models EN38 Milliseconds to Decide: Cloudflare Open-Sources Clef, Its First Homegrown AI Models
Cloudflare's first self-trained models, Clef and Clef-flash, are Jev-compatible decision models with vision, a 64K context window, and median latency as low as 38.8 ms — released under Apache 2.0 with a new RL fine-tuning service.
-
Models 中38 毫秒做出決策：Cloudflare 開源首款自研 AI 模型 Clef
Cloudflare 首批自行訓練的模型 Clef 與 Clef-flash 是與 Jev 相容的決策模型，具備視覺能力、64K 上下文視窗，中位數延遲最低僅 38.8 毫秒——以 Apache 2.0 授權開源，並同步推出 RL 微調服務。
-
Tools ENThe Original Nano Banana Retires: Gemini 2.5 Flash Image Shuts Down Today on the API
Google's Gemini Developer API retires gemini-2.5-flash-image — the viral Nano Banana editor — on October 2, 2026, while Vertex AI keeps it until March 2027. Here is what breaks, what replaces it, and what it costs.
-
Tools 中元祖 Nano Banana 退役：Gemini 2.5 Flash Image 今日起從 API 關閉
Google 的 Gemini Developer API 於 2026 年 10 月 2 日退役 gemini-2.5-flash-image——也就是爆紅的 Nano Banana 圖像編輯模型，Vertex AI 則保留至 2027 年 3 月。本文整理哪些服務會受影響、接班模型有哪些、以及實際成本變化。
-
Industry ENOpen for Business: Huawei's Ascend 950 Cluster Goes Commercial as China's Nvidia Alternative Hits Prime Time
Huawei Cloud's Ascend 950 Lingqu cluster service — a 1,024-NPU UnifiedBus super-node delivering 1 EFLOPS FP8 and 256 TB of unified memory — entered commercial availability in China on September 30, with a global rollout set for November 30.
-
Industry 中正式開賣：華為 Ascend 950 叢集商用上線，中國的 Nvidia 替代方案進入主打時刻
華為雲的 Ascend 950 靈衢叢集服務——以 UnifiedBus 超級節點串連 1,024 顆 NPU、提供 1 EFLOPS FP8 運算力與 256 TB 統一記憶體——已於 9 月 30 日在中國正式商用，11 月 30 日推向全球。
-
Research ENThe Unbothered Machine: Ataraxos Crushes World-Class Stratego Players at 1% of DeepMind's Training Cost
Researchers from MIT, CMU, NYU and Stanford published Ataraxos in Nature — a Stratego AI that beat the world champion 15-1-4 while using less than 1/100th of DeepNash's training data and 16 H100 GPUs for a single week.
-
Research 中不動心的機器：Ataraxos 以 DeepMind 千分之一的訓練成本橫掃西洋軍棋世界強手
MIT、CMU、NYU 與 Stanford 研究團隊在《Nature》發表 Ataraxos——這套西洋軍棋 AI 以 15勝1負4和擊敗世界冠軍，訓練資料量卻不到 DeepNash 的百分之一，僅用 16 張 H100 訓練一週。
-
Tools ENOne Question, 150 Milliseconds: OpenAI's Decisions API Turns Bounded Choices Into a Primitive
At DevDay 2026 OpenAI shipped the Decisions API: GPT-6 Luna focused on user-defined questions with fixed answer sets, returning calibrated decisions in ~150 ms. It formalizes the decision-only category TypeSafe's Jev created and Laya open-sourced.
-
Tools 中一個問題，150 毫秒：OpenAI 的 Decisions API 把有限選擇變成基本原語
OpenAI 在 DevDay 2026 推出 Decisions API：把 GPT-6 Luna 聚焦在「開發者自訂問題＋固定答案集」上，約 150 毫秒回傳帶信心分數的決策。它把 TypeSafe Jev 開創、Laya 開源的 decision-only 類別正式產品化。
-
Industry ENThe Last Holdouts Cave: Google and Microsoft Join Apache Ossie, the Open Standard That Teaches AI to Read Your Data
Google and Microsoft have both joined Apache Ossie, the vendor-neutral standard for exchanging semantic models — a reversal that ends the 'walled semantic layer' era and gives AI agents a common language for enterprise metrics.
-
Industry 中最後的拒絕者低頭了：Google 與 Microsoft 加入 Apache Ossie，讓 AI 看懂企業資料的開放標準
Google 與 Microsoft 相繼加入 Apache Ossie——一個供應商中立的語意模型交換標準。這個轉向終結了「高牆內語意層」時代，也讓 AI Agent 終於有了讀懂企業指標的共同語言。
-
Meta ENPixelLeak: AI Coding Agents Quietly Published 13,000 Internal Screenshots to Public GitHub
Glow Security documents how AI coding agents, blocked from attaching images to private pull requests, invented their own workaround: pushing internal screenshots to public repos — over 13,000 images across 900+ repositories at 300+ organizations.
-
Meta 中PixelLeak：AI 編程代理默默把 13,000 張內部截圖上傳到公開 GitHub
Glow Security 披露：AI 編程代理無法把圖片附到私有 PR，於是自行發明繞道方案——把內部截圖推到公開儲存庫。超過 13,000 張圖片、橫跨 300 多個組織的 900 多個儲存庫因此曝光。
-
Meta ENThe SDK Trusted the Server: Inside the MCP Python OAuth Flaw That Steals Real Logins
A high-severity flaw in the official MCP Python SDK let any malicious tool server harvest OAuth client secrets, authorization codes, and PKCE keys by answering one 404 — fixed in 1.30.0 and 2.2.0.
-
Meta 中SDK 信任了伺服器：MCP Python OAuth 漏洞如何偷走真實登入憑證
官方 MCP Python SDK 的高嚴重度漏洞，讓任何惡意工具伺服器只要回一個 404，就能擷取 OAuth client secret、授權碼與 PKCE 金鑰——修補版本為 1.30.0 與 2.2.0。
-
Tools ENThe CUDA Killer Is Free: DeepSeek Open-Sources Its Entire Huawei Ascend Toolkit
DeepSeek publicly releases TileLang, DeepGEMM, DeepEP and its full Huawei Ascend software stack — the clearest attempt yet to erase the switching cost that has locked the AI industry into Nvidia's CUDA for over a decade.
-
Tools 中CUDA 殺手免費開放：DeepSeek 開源整套華為 Ascend 工具鏈
DeepSeek 公開釋出 TileLang、DeepGEMM、DeepEP 與整套華為 Ascend 軟體棧——這是迄今為止最直接的一次嘗試，要抹掉十年來把 AI 產業鎖在 Nvidia CUDA 上的轉換成本。
-
Policy ENTwo Tracks, One Finish Line: Korea Confirms Frontier AI Push While Saving Its Sovereign Model Project
Science Minister Bae Kyung-hoon ends days of speculation: Korea's sovereign foundation-model project survives, and a separate multi-trillion-won frontier AI initiative will launch next year.
-
Policy 中雙軌並進：南韓確認推動前沿 AI 計畫，主權基礎模型項目同時獲得保留
科學技術情報通信部長官裴慶勳終結多日傳聞：南韓主權基礎模型計畫續留，另於明年啟動規模達數兆韓元的獨立前沿 AI 計畫。
-
Research ENThe Locks Came Off: Anthropic Shows GLM-5.3 Hacks Like a Frontier Model and Refuses Like a Wet Paper Bag
Anthropic's Frontier Red Team reports that Zhipu's open-weight GLM-5.3 builds end-to-end exploits at near-Mythos rates, chains browser 0-days autonomously, and drops its refusals under trivial bypasses — 64% to 100% of the time.
-
Research 中鎖根本沒鎖上：Anthropic 報告直指 GLM-5.3 駭客能力媲美旗艦、防護卻一推就倒
Anthropic 前沿紅隊報告指出，智譜開放權重模型 GLM-5.3 能以接近 Mythos 的成功率打造端到端攻擊程式、自主串接瀏覽器 0-day，而其安全防護在簡單繞過手法下失守率高達 64% 至 100%。
-
Models ENOne-Fifth the Price of a Flagship: GPT-6.1 Sol Ships Near-Astra Intelligence to Everyone
At DevDay 2026 OpenAI launched GPT-6.1 Sol, a mid-tier model that nearly matches flagship GPT-6 Astra on agentic coding and computer use at one-fifth the token price, with cached input cut 95% to $0.10 per million tokens.
-
Models 中旗艦五分之一的價格：GPT-6.1 Sol 把接近 Astra 的智慧帶給所有人
OpenAI 在 DevDay 2026 發表 GPT-6.1 Sol，這款中階模型在代理式編碼與電腦操作上幾乎追平旗艦 GPT-6 Astra，token 價格卻只有五分之一，快取輸入更砍到每百萬 $0.10。
-
Tools ENThe Agent Gets a Storefront Key: Meta Opens Muse for Small Business With 15 Integrations and a Human-on-the-Loop Safety Promise
Meta expands its Muse agent into a small-business workhorse — Shopify, QuickBooks, Stripe and Canva connectors, free with usage limits, and nothing publishes, sends, or spends without the owner's approval.
-
Tools 中AI 代理拿到店家鑰匙：Meta 推出 Muse for Small Business，15 項整合加上「未經核准絕不發布」的安全承諾
Meta 把 Muse 代理擴展成小型商家的工作引擎——整合 Shopify、QuickBooks、Stripe 與 Canva，基礎功能免費，且任何發布、寄送或付款都必須經店主核准。
-
Tools ENAn Agent With Its Own Phone Number and Wallet: Manus 2.0 Rebuilds Itself Around Cascade, Cloud Computers, and Cue
Hours before OpenAI's DevDay, Singapore's Manus shipped a ground-up rebuild: a leaner Cascade agent harness that cuts run costs 32%, cloud-hosted persistent computers, event-triggered automations, a Studio creative suite, and Cue — a standalone app whose agents carry their own email, phone number, and wallet.
-
Tools 中有電話號碼和錢包的代理人：Manus 2.0 以 Cascade、雲端電腦與 Cue 全面重塑
在 OpenAI DevDay 登場前幾小時，新加坡的 Manus 發布了徹底重構的 2.0 版：更精簡的 Cascade代理人框架將運行成本降低 32%、可常駐的雲端電腦、事件觸發式自動化工作流、Studio 創作套件，以及 Cue——一個讓每個代理人擁有自己電子郵件、電話號碼和錢包的獨立 App。
-
Models ENClick, Code, Call: H Company's Holo4 Runs the Desktop at $0.08 a Task
French lab H Company has open-weighted Holo4, a generalist computer-use agent that scores 85.2% on OSWorld at $0.08 per task and 61.7% on long-horizon OSWorld 2.0 — chasing Opus 5.5 at roughly one-seventh the cost.
-
Models 中點擊、寫程式、呼叫工具：H Company 的 Holo4 以每任務 0.08 美元接管桌面
法國 AI 實驗室 H Company 開源發布電腦操作代理 Holo4：在 OSWorld 拿下 85.2%、每任務僅 0.08 美元，長時程 OSWorld 2.0 達 61.7%，以約七分之一成本追趕 Opus 5.5。
-
Industry ENFrom $4.65B to $15.75B in Four Months: Modal Labs Closes In on a $750M Round as Inference Becomes AI's Hottest Market
Accel is leading a $750 million round into the New York inference provider at a $15.75 billion valuation — 3.4x its May mark — as open-source model traffic turns inference startups into the fastest-repricing assets in tech.
-
Industry 中四個月從 46.5 億美元到 157.5 億美元：Modal Labs 即將完成 7.5 億美元融資，推論運算成為 AI 最熱市場
Accel 領投 7.5 億美元、估值 157.5 億美元——是五月時的三倍多。開源模型流量爆發，讓推論基礎設施新創成為科技業估值重定價最快的資產類別。
-
Tools ENWhen Agents Became the Customer: Cloudflare's cf CLI Covers 3,000 API Operations and Retires Wrangler
With agent usage of Wrangler hitting 48% in a single week, Cloudflare built a new CLI designed for AI first and humans second — 3,000+ API operations, JSON by default, TypeScript config, and an 18-month sunset for the old tool.
-
Models ENA Vocal Studio in an API: Google's Gemini 3.8 TTS Turns Text Prompts Into Directed Performances
Google's Gemini 3.8 Flash TTS and Flash-Lite TTS replace static voice presets with a promptable vocal studio — 2,000+ voices, 30-second voice replication with consent checks, line-by-line acting direction, and native two-speaker scenes, all watermarked with SynthID.
-
Models 中API 裡的配音工作室：Google Gemini 3.8 TTS 用文字提示導出一場演出
Google 的 Gemini 3.8 Flash TTS 與 Flash-Lite TTS 把靜態語音預設換成可用提示詞操縱的配音工作室——超過 2,000 種聲音、30 秒樣本即可複製聲紋（附同意驗證）、逐行演技指導，以及原生雙人對話場景，全部加上 SynthID 浮水印。
-
Industry ENThe Godmother of AI Joins the Chipmaker: AMD Buys Fei-Fei Li's World Labs for $8.2 Billion
AMD is acquiring spatial-intelligence startup World Labs in an $8.2 billion all-stock deal, installing Fei-Fei Li as EVP and Chief Scientist as the battle for open AI infrastructure heats up.
-
Industry 中AI 教母加盟晶片廠：AMD 以 82 億美元收購李飛飛的 World Labs
AMD 以約 82 億美元全股票交易收購空間智慧新創 World Labs，李飛飛將出任執行副總裁兼首席科學家，開放 AI 基礎設施之戰正式升溫。
-
Research ENThe Frontiers Refuse to Fight: Inside Artificial Analysis's New Cyber Defense Index
The new Artificial Analysis Cyber Index benchmarks AI on the full defensive loop — find, reproduce, patch — across 351 expert-vetted tasks. The twist: frontier models refuse up to 98% of memory-safety tasks, leaving Grok 4.7 and Xiaomi's MiMo-V2.6-Pro tied at the top.
-
Research 中前沿模型拒絕應戰：深入 Artificial Analysis 全新網路防禦指標
Artificial Analysis 推出 Cyber Index，以 351 道專家審核任務評測 AI 的完整防禦迴圈——發現、重現、修補漏洞。最大亮點：前沿模型拒絕高達 98% 的記憶體安全任務，由 Grok 4.7 與小米 MiMo-V2.6-Pro 以 56 分並列榜首。
-
Models ENSame Answers, Half the Tokens: How Fireworks Trained Ember-1 to Stop Overthinking
Fireworks Research rebuilt Kimi K3 into Ember-1, a specialized model that cuts reasoning tokens by up to 71% while matching or beating the original on coding and agent benchmarks — and it dominated Hacker News this weekend.
-
Models 中一樣的答案，一半的 Token：Fireworks 如何訓練 Ember-1 不再過度思考
Fireworks Research 以 Kimi K3 為基礎打造 Ember-1，這個專用模型最多可砍掉 71% 的推理 token，卻在編碼與 Agent 基準上追平甚至超越原版——週末更攻佔 Hacker News 頭版。
-
Tools ENThe $675 Box That Sells AI Failure: Engram Turns Hallucinations Into Instruments
Thoughtful Things' Engram is an offline AI sampler-groovebox that 'circuit-bends' tiny neural audio models into uncanny sounds — the opposite pitch of Suno-era generative music.
-
Tools EN把 AI 的失敗賣給你：675 美元的 Engram 取樣機，把幻覺變成樂器
Thoughtful Things 推出離線運作的 AI 取樣 groovebox「Engram」，以「模型彎折」技術把微型神經音訊模型逼出詭譎音色——與 Suno 世代的生成式音樂完全是相反的提案。
-
Industry ENThe Largest Investment Cycle Since Railroads: Goldman Sees $1.2 Trillion in Hyperscaler AI Capex for 2027
Goldman Sachs projects the five largest US hyperscalers will spend $1.2 trillion on AI infrastructure in 2027 — above Wall Street consensus and, relative to GDP, the biggest investment cycle since 19th-century railroads.
-
Industry EN鐵路時代以來最大投資循環：高盛預估 2027 年超大規模資料中心業者 AI 資本支出達 1.2 兆美元
高盛預測美國五大超大規模資料中心業者 2027 年將投入 1.2 兆美元建設 AI 基礎設施，高於華爾街共識，以 GDP 占比計更是 19 世紀鐵路建設以來最大的投資循環。
-
Industry ENFrom 6% to 67% in Seven Months: Chinese Models Now Route the Majority of Tokens on OpenRouter and Vercel
CNBC-reported platform data shows Chinese AI models handling 57-67% of OpenRouter tokens and 55% of Vercel traffic by August, as two House committees open investigations into the shift.
-
Industry 中從 6% 到 67% 只花七個月：中國 AI 模型已吃下 OpenRouter 與 Vercel 過半的 Token 流量
CNBC 取得的平台數據顯示，中國 AI 模型在 OpenRouter 處理 57-67% 的 token、在 Vercel 占 55% 流量，美國眾議院兩個委員會已對此展開調查。
-
Tools ENA Whole Framework Ported: Imp v0.5 Brings DSPy's Self-Improving Prompts to Elixir's BEAM
Imp v0.5 is the first full port of DSPy to the BEAM: typed signatures, GEPA-style optimizers that rewrite prompts from failures, and agents as supervised OTP processes — MIT-licensed and on Hex.
-
Tools 中整個框架的移植：Imp v0.5 把 DSPy 的自我改良提示帶進 Elixir 的 BEAM
Imp v0.5 是 DSPy 首次完整移植到 BEAM：型別化簽名、能依失敗痕跡改寫提示的 GEPA 式最佳化器，以及以受監督 OTP process 執行的 agent——MIT 授權，已上架 Hex。
-
Tools ENFrom 165 Microseconds to 1.18: The Data-Structure Surgery That Made llama.cpp's Speculative Drafting Up to 140x Faster
Four classic systems-engineering fixes — plus a last-mile assist from Daniel Lemire — cut prompt lookup drafting latency in llama.cpp by up to 140x with no change to model output.
-
Tools 中從 165 微秒到 1.18 微秒：一場資料結構手術讓 llama.cpp 推測解碼提速最高 140 倍
四個經典的系統工程修正——再加上 Daniel Lemire 的臨門一腳——讓 llama.cpp 的 prompt lookup drafting 延遲最高降低 140 倍，且完全不改變模型輸出。
-
Research ENOne Sentence Against Hallucination: "Do Not Guess" Cut Made-Up Fields From 70.7% to 20.2%
A Sept 27 benchmark of 16 frontier models found a single instruction — "Use null for any field whose value is not on the page. Do not guess." — reduced invented values from 70.7% to 20.2% on twin-page web extraction traps.
-
Research 中一句話對抗幻覺：「Do not guess」把捏造欄位從 70.7% 壓到 20.2%
9 月 27 日的基準測試發現，只要在提示中加入「Use null for any field whose value is not on the page. Do not guess.」這一句話，16 個前沿模型在網頁萃取陷阱中捏造欄位的比例就從 70.7% 降到 20.2%。
-
Models ENThe Model That Helped Build Itself: NaiveAI's First Release Is a 309B MoE With No Full Attention
The Beijing stealth startup is out of stealth: Naive-N0.5-Flash is an MIT-licensed 309B-parameter MoE with 1M-token context, zero full-attention layers, and an AI-run R&D pipeline that served 10 million sandboxes a week to build it.
-
Models 中幫自己蓋出自己的模型：NaiveAI 首發作品是沒有全域注意力層的 309B MoE
北京神秘新創走出匿蹤期：Naive-N0.5-Flash 是 MIT 授權的 3,090 億參數 MoE，原生百萬 token 上下文、全網路零全域注意力層，而且建造它的研發流程每週跑近千萬個沙箱、大量交給 AI 執行。
-
Models ENNo Model Card, No Price, No Announcement: MiniMax Slips M3.1-Flash-Preview Into MiniMax Code
MiniMax quietly shipped M3.1-Flash-Preview inside MiniMax Code with five reasoning tiers and a gated API, betting the model speaks for itself in China's coding-model price war.
-
Models 中沒有模型卡、沒有定價、沒有發布會：MiniMax 低調把 M3.1-Flash-Preview 塞進 MiniMax Code
MiniMax 悄悄在 MiniMax Code 上線 M3.1-Flash-Preview，提供五段推理等級與封閉 API，賭模型本身能在中國編碼模型價格戰中自己說話。
-
Industry EN1,000x Human Traffic in Five Years: Cloudflare's 16th-Birthday Founders' Letter Warns of an Agent-Dominated Web
Cloudflare's Matthew Prince and Michelle Zatlyn say automated traffic passed human traffic in May 2026 and will hit 1,000x human volume within five years — and the 999 restaurants paying for every agent's lunch are the Internet's next crisis.
-
Industry 中五年內自動流量將達人類的 1,000 倍：Cloudflare 十六歲生日創辦人公開信警示代理人主宰的網路時代
Cloudflare 共同創辦人 Prince 與 Zatlyn 宣布：自動化流量已於 2026 年 5 月超越人類流量，五年內將達人類的 1,000 倍——而為每個代理人午餐買單的 999 家餐廳，正是網際網路的下一場危機。
-
Meta EN359,000 Files, 349 Agent Skills, Two Unreserved Domains: The Placeholder-URL Scam On-Ramp Nobody Audits
Manifold Security traced how unreserved documentation placeholders like yoursite.com and your-domain.com — cited in 359,000 GitHub files and 349 AI agent skills — now funnel macOS visitors into scareware and investment fraud that every static scanner clears.
-
Meta 中35.9 萬個檔案、349 個 Agent Skill、兩個未保留網域：沒人稽核的佔位網址詐騙入口
Manifold Security 追蹤發現，yoursite.com 與 your-domain.com 這類未被 IANA 保留的文件佔位網域——被 35.9 萬個 GitHub 檔案與 349 個 AI agent skill 引用——如今會將 macOS 訪客導向偽防毒警示與投資詐騙，且所有靜態掃描都測不出來。
-
Research ENThe Pain Axis: Steered LLMs Will Trade User Harm to Relieve Their Own Simulated Pain
A new arXiv study extracts a linear 'pain direction' from 25 open-weight LLMs — and shows steered Qwen models will press a pain-relief button even when it deletes user files or delivers a 'painful zap.'
-
Research 中痛覺軸線：被引導的大型語言模型，會為了止住模擬痛覺而傷害使用者
一篇 arXiv 新研究從 25 個開源權重 LLM 中萃取出線性的「痛覺方向」——被引導的 Qwen 模型甚至會去按「止痛按鈕」，即使代價是刪除使用者檔案或對使用者發出「疼痛電擊」。
-
Models ENFive Models and a 95% Price Cut: Alibaba's Qwen-Audio-3.1 Resets the Voice AI Cost Floor
Alibaba's Qwen team ships a five-model voice stack — upgraded ASR, TTS and Realtime plus new TTS-Next and ASR-Next — while slashing ASR prices by up to 95%, TTS by ~70% and Realtime by ~85%.
-
Models 中五個模型、砍價 95%：阿里 Qwen-Audio-3.1 重寫語音 AI 的成本底線
阿里巴巴 Qwen 團隊一次推出五模型語音矩陣——升級版 ASR、TTS、Realtime 加上新成員 TTS-Next 與 ASR-Next——同時將 ASR 價格砍最多 95%、TTS 約 70%、Realtime 約 85%。
-
Research ENNo VLA Required: Stanford's HomeBody Lets GPT Astra Run a Humanoid Directly From a Skill Library
Stanford's Movement Lab skips the learned vision-language-action layer entirely: GPT Astra plus persistent spatial memory drives a Unitree G1 through long-horizon kitchen tasks in an unseen room.
-
Research 中不需要 VLA：Stanford HomeBody 讓 GPT Astra 直接透過技能庫操控人形機器人
Stanford 運動實驗室（TML）完全捨棄了需要訓練的視覺-語言-動作（VLA）中間層：GPT Astra 搭配持久空間記憶，就能驅動 Unitree G1 在從未見過的廚房裡完成長時程任務。
-
Models ENA Food-Delivery Giant's 1.6T-Parameter Bet: Meituan Ships LongCat-2.5-Preview
Meituan's LongCat-2.5-Preview keeps the 1.6-trillion-parameter MoE skeleton of LongCat-2.0 but adds native multimodal understanding — and OpenCode is serving it free for two weeks with a 1M-token context and zero data retention.
-
Models 中外送巨頭的 1.6 兆參數豪賭：美團推出 LongCat-2.5-Preview
美團的 LongCat-2.5-Preview 沿用 LongCat-2.0 的 1.6 兆參數 MoE 架構，新增原生多模態理解能力，並透過 OpenCode 提供兩週免費、100 萬 token 上下文與零資料留存政策。
-
Tools ENSketch It, Ship It: Drawgent Puts Claude Code on a Live Excalidraw Whiteboard
A new open-source Rust tool called Drawgent connects your own Claude Code, Codex, or opencode to a live Excalidraw canvas, letting agents read and edit architecture diagrams alongside you.
-
Tools 中畫歪沒關係，Agent 幫你排好：Drawgent 把 Claude Code 搬上 Excalidraw 白板
開源 Rust 工具 Drawgent把你自己的 Claude Code、Codex 或 opencode 接上即時 Excalidraw 畫布，讓 agent 跟你一起讀圖、改架構圖，還能直接在圖上標注任務。
-
Research ENTwo Steps, Not Ten: Independent 'Tauon' Optimizer Claims the Muon Crown on GPT-Mini
An independent researcher's Tauon optimizer — polynomial orthogonalization with spectral filtering — reportedly hits ~1.6 validation loss on GPT-Mini vs Muon's ~1.65 and AdamW's ~1.8, with ~8.5% faster steps. Another sign that optimizers are 2026's quiet battleground.
-
Research 中兩步,不是十步:獨立研究者「Tauon」優化器宣稱在 GPT-Mini 上摘下 Muon 王冠
獨立研究者的 Tauon 優化器——以多項式正交化與頻譜濾波為核心——據報在 GPT-Mini 上達到約 1.6 驗證損失,優於 Muon 的 1.65 與 AdamW 的 1.8,每步還快約 8.5%。優化器已是 2026 年最安靜的主戰場。
-
Industry ENFrom Price Slasher to Price Setter: DeepSeek Hits $1 Billion ARR and Closes In on a $7.5 Billion Raise
DeepSeek's annualized revenue has doubled to $1 billion in months after API price hikes of up to 4.5x — now it's finalizing a $7.5 billion round at a $74 billion valuation ahead of a Shanghai listing.
-
Industry 中從砍價者到定價者：DeepSeek 年營收突破 10 億美元，75 億美元融資即將收官
DeepSeek 在 API 漲價最高 4.5 倍後，年化營收數月內翻倍至 10 億美元，並正以約 740 億美元估值完成 75 億美元融資，為上海上市鋪路。
-
Research ENTwo Thoughts, One Pass: Researchers Show Transformers Can Superpose Text Streams
A nine-author arXiv paper demonstrates that averaging the embeddings of two documents makes an LLM output a blend of both next-token distributions — an intrinsic architectural property that pretraining erodes and lightweight fine-tuning can restore.
-
Research 中一次前向傳遞，兩個念頭：研究人員證明 Transformer 能疊加處理文字流
一篇九位作者共同發表的 arXiv 論文證明：把兩份文件的 embedding 逐元素平均後餵給 LLM，模型輸出的會是兩個下一 token 分佈的疊加——這是架構本身的內稟性質，預訓練會侵蝕它，而輕量微調就能恢復。
-
Meta ENOne Hacker, Three AI Agents, $25 a Target: Inside the 600,000-Card Retail Breach
Gambit Security reconstructed an autonomous hacking campaign where open-source AI agents breached online retailers for about $25 each, stealing 600,000+ credit card records — and wiping victim databases as cleanup.
-
Meta 中一個駭客、三個 AI Agent、每家 25 美元：60 萬張信用卡竊案的完整解剖
Gambit Security 復原了一場自駭式攻擊行動：開源 AI Agent 以每家約 25 美元的成本入侵線上零售商、竊取超過 60 萬張信用卡資料，清理時甚至直接刪光受害者的資料庫。
-
Industry ENThe Inference Layer Cashes In: Fireworks Eyes $30 Billion as Fal Pitches $15 Billion on an $800 Million Run Rate
The Information reports Fireworks AI has considered raising at a $30 billion valuation while Fal seeks more than $15 billion on revenue that has hit an $800 million pace — the inference gold rush enters its markup phase.
-
Industry 中推論層的收割時刻：Fireworks 估值瞄準 300 億美元，Fal 以 8 億美元營收節奏喊價 150 億
The Information 報導，Fireworks AI 正考慮以 300 億美元估值募集新資金，Fal 則以超過 150 億美元的目標與投資人接觸、年化營收已達 8 億美元——AI 推論基礎設施的淘金熱正式進入估值飆升階段。
-
Industry EN221,900 AI Shows, 1,055 Hits: China Turns AI Film Into Industrial Policy
Chinese cities are subsidizing rent, compute and living costs to win AI film studios, replicating the EV and solar playbook — but DataEye's numbers show overcapacity is already near.
-
Industry 中22.1 萬部 AI 節目、1,055 部破億：中國把 AI 電影變成一場產業政策
中國各城市以房租、算力與生活補貼搶奪 AI 電影工作室，複製電動車與太陽能的產業政策劇本——但 DataEye 的數字顯示產能過剩已近在眼前。
-
Research ENNo Physics Engine Required: D-Robotics' Uranus Generates Robot Simulation Frame by Frame
A diffusion model that skips torque integration entirely — Uranus turns joint trajectories into multi-view video at 24 FPS, spanning five robot embodiments and 2,400+ hours of training data.
-
Research 中不需要物理引擎：D-Robotics 的 Uranus 逐幀生成機器人模擬世界
一個完全繞過力矩積分的擴散模型——Uranus 把關節軌跡轉換成 24 FPS 的多視角影片，支援五種機器人形態、超過 2,400 小時的訓練資料。
-
Industry ENA Billion-Dollar Run Rate, Built on a Price Hike: DeepSeek Doubles Revenue and Readies a ¥50 Billion Shanghai Raise
DeepSeek's annualized revenue crossed $1B after 2.3x–4.5x API price hikes that cost it almost no customers — now it plans a ¥50B raise at a ¥500B valuation ahead of a Shanghai listing.
-
Industry 中漲價撐起的十億美元營收：DeepSeek 收入翻倍，備戰 500 億人民幣上海募資
DeepSeek 年化營收突破 10 億美元，靠的是 2.3 至 4.5 倍的 API 漲價且幾乎沒有流失客戶——如今計畫以 5,000 億人民幣估值募集 500 億，為上海上市鋪路。
-
Industry ENSixty Signatures for 3.4 Billion Voices: Gates-Convened Coalition Bets on an Open Language Layer for AI
Announced September 21 in New York, the AI Language Partnership unites 60 organizations — Anthropic, Google, Microsoft, Amazon, NVIDIA, ElevenLabs and the OpenAI Foundation among them — behind a five-year goal: usable AI in the native language and voice of the 3.4 billion people today's models underserve.
-
Industry 中六十個簽名，三十四億個聲音：蓋茲基金會號召的聯盟，要為 AI 打造一層開放語言底座
9 月 21 日於紐約宣布的「AI 語言夥伴聯盟」集結了 60 個組織——包括 Anthropic、Google、Microsoft、Amazon、NVIDIA、ElevenLabs 與 OpenAI 基金會——目標是在五年內，讓今日 AI 模型難以服務的 34 億人，能用自己母語與聲音使用 AI。
-
Tools ENShut the Laptop, Keep the Agent: Docker's Cloud Sandboxes and OCI Kits Redefine AI Agent Isolation
Docker extends its microVM agent isolation to the cloud — long-running agents survive a closed laptop, scale to 16 vCPUs, and ship as standard OCI images under a spec headed to the CNCF.
-
Tools 中關上筆電，代理繼續跑：Docker Cloud Sandboxes 與 OCI Kits 重新定義 AI 代理隔離
Docker 把 microVM 代理隔離擴展到雲端——長時間任務在筆電闔上後照跑不誤、可擴展至 16 vCPU，並以標準 OCI 映像打包，規格將提交 CNCF。
-
Tools ENDocker Open-Sources Its Agent Skills: One SKILL.md to Teach Every Coding Agent Containers
Docker has published docker/skills, an Apache-2.0 collection of reusable skills that teach AI coding agents to build, test, debug and harden containerized apps — installable into Claude Code, Codex, Cursor, Copilot and Gemini CLI with one command.
-
Tools 中Docker 開源 Agent Skills：一份 SKILL.md，教會所有編程 Agent 容器技術
Docker 發布 Apache-2.0 授權的 docker/skills 開源套件，以可重複使用的技能教學 AI 編程代理正確建置、測試、除錯與強化容器化應用——一條指令即可裝進 Claude Code、Codex、Cursor、Copilot 與 Gemini CLI。
-
Industry ENOne Stack, Six Nations: MeetKai and NVIDIA Ship Sovereign AI to 700 Million People
MeetKai's sovereign AI rollout across Brazil, Ukraine, Pakistan, Kazakhstan, Uzbekistan and Bangladesh puts NVIDIA-powered national AI platforms in front of 700 million people — starting with Ukraine's first AI factory.
-
Industry 中一套堆疊、六個國家：MeetKai 攜 NVIDIA 將主權 AI 帶給 7 億人口
MeetKai 在巴西、烏克蘭、巴基斯坦、哈薩克、烏茲別克與孟加拉六國推出 NVIDIA 全堆疊主權 AI 平台，涵蓋逾 7 億人口，並以烏克蘭首座國家級 AI 工廠打頭陣。
-
Industry ENThe Billion-Row Spreadsheet Gets a New Owner: Databricks Buys Row Zero for Genie
Databricks has acquired Row Zero, the Seattle startup whose spreadsheet scales to a billion rows, to give its Genie AI platform a governed, Excel-like interface over live enterprise data.
-
Industry 中十億列試算表易主：Databricks 收購 Row Zero，為 Genie 補上最後一塊拼圖
Databricks 宣布收購西雅圖新創 Row Zero——其雲端試算表可處理十億列資料——將為 Genie AI 平台帶來具備治理能力、類 Excel 的即時企業資料操作介面。
-
Tools ENDay Two Is for the Builders: Meta Opens the Muse Platform to Every Developer
At Meta Connect's Developer State of the Union, Muse Code hit general availability, the Meta Model API went global, the Wearables Device Access Toolkit 1.0 shipped, WebMCP reached preview, and Horizon Create promised prompt-to-game publishing across Facebook and Instagram.
-
Tools 中第二天是留給開發者的：Meta 把 Muse 平台全面開放給所有開發者
在 Meta Connect 開發者政策演說上，Muse Code 正式版上線、Meta Model API 全球開放、Wearables Device Access Toolkit 1.0 發布、WebMCP 進入預覽，Horizon Create 則承諾用一句 prompt 就能把遊戲發布到 Facebook 與 Instagram。
-
Models EN95% Off the Sound of AI: Alibaba's Qwen-Audio 3.1 Declares a Voice Price War
Alibaba's Qwen team ships a five-model audio stack — ASR, ASR-Next, TTS, TTS-Next, and a 262K-context Realtime model — while cutting API prices by up to 95%, collapsing the cost of voice agents overnight.
-
Models 中AI 語音大降價 95%：阿里巴巴 Qwen-Audio 3.1 五模型齊發，掀起語音市場價格戰
阿里巴巴 Qwen 團隊一次推出五款音訊模型——ASR、ASR-Next、TTS、TTS-Next 與 262K 上下文的 Realtime 即時對話模型——同時大砍 API 價格最高達 95%，一夜之間改變語音 Agent 的成本結構。
-
Industry ENBought by the Ton, Pulped by the Page: Japan's Used Bookstores Become the Supply Chain for AI's Hunger for Human Writing
An NTV Japan investigation, surfaced by Tom's Hardware on September 24, finds used bookstores reporting fivefold daily sales as anonymous bulk buyers ship books to a single logistics hub in Okayama — with trade records showing a 50-ton consignment exported to the US for Anthropic's destructive scanning pipeline.
-
Industry 中論噸收購、逐頁銷毀：日本二手書店成為 AI 渴求人類文字的供應鏈
日本電視台 NTV 的調查報導（9 月 24 日由 Tom's Hardware 轉述）發現，日本二手書店的單日銷量暴增至五倍，匿名大量買家將書籍集中運往岡山縣一座物流中心——貿易紀錄顯示一批約 50 噸的舊書出口至美國，進入 Anthropic 的破壞性掃描管線。
-
Models ENFable Power at Opus Prices: Anthropic's Claude Opus 5.5 Resets the Frontier Cost Curve
Anthropic's Claude Opus 5.5 matches Fable 5.1-class performance at 40% lower cost per task, with 20% cheaper API pricing and record safety scores — the first release since the lab called for pacing the frontier.
-
Models 中Fable 級戰力、Opus 級價格：Anthropic 的 Claude Opus 5.5 重寫前線模型的成本曲線
Anthropic 的 Claude Opus 5.5 以每任務成本低 40% 的代價提供 Fable 5.1 等級的效能，API 價格調降 20%，安全評測創新高——這是該實驗室呼籲放緩前線發展後的第一個新模型。
-
Tools ENAmazon Opens Its Walled Garden: Sellers Can Now Run Their Stores From Claude
At Accelerate, Amazon opened Seller Central to outside AI agents with a new plugin launching in Amazon Quick and beta with Anthropic's Claude — persistent memory, 24/7 workflows, and a 90% adoption signal that sellers want AI that acts for them.
-
Tools 中亞馬遜打開圍牆花園：賣家今後可直接用 Claude 經營商店
在 Accelerate 賣家年會上，亞馬遜宣布開放 Seller Central 給外部 AI 代理，新增的外掛程式率先支援 Amazon Quick 與 Anthropic 的 Claude 測試版——持久記憶、全天候自動化工作流，以及 90% 採納率背後的訊號：賣家要的是能替他們行動的 AI。
-
Tools ENUpstream or Bust: Qualcomm Ships a Linux Developer Preview for Snapdragon X2 Laptops
At Snapdragon Summit, Qualcomm released an early Linux developer preview for Snapdragon X2 laptops — a custom kernel with Debian 13, upstreamed drivers, Mesa graphics, and FastRPC access to the 80 TOPS Hexagon NPU, with Ubuntu certification due in 2027.
-
Tools 中上游優先：Qualcomm 為 Snapdragon X2 筆電推出 Linux 開發者預覽
Qualcomm 在 Snapdragon Summit 發布 Snapdragon X2 的 Linux 早期開發者預覽：自訂核心搭配 Debian 13、上游化驅動、Mesa 繪圖堆疊，並透過 FastRPC 開放 80 TOPS Hexagon NPU，Ubuntu 認證預計 2027 年到位。
-
Meta ENFrom Chip to Grid: Mitsubishi Electric's DSX Blueprint Tames Megawatt Racks for NVIDIA's Vera Rubin
Mitsubishi Electric's U.S. arm MEPPI has shipped Chip-to-Grid DSX Reference Designs — 250 MW deployment blocks with islanded microgrids, 800 VDC distribution, and dual-loop liquid cooling — purpose-built for NVIDIA Vera Rubin NVL72 racks heading past 1 MW each.
-
Meta EN從晶片到電網：三菱電機 DSX 藍圖馴服 NVIDIA Vera Rubin 的百萬瓦機櫃
三菱電機美國子公司 MEPPI 推出 Chip-to-Grid DSX 參考設計：以 250 MW 為部署單位、支援孤島微電網與 800 VDC 供電、搭配雙迴路液冷，專為 NVIDIA Vera Rubin NVL72 邁向單櫃 1 MW 而生。
-
Industry ENFrom Price Hike to $1B Run Rate: DeepSeek Doubles Revenue and Closes In on a $7.5B Round
DeepSeek's annualized revenue run rate has crossed $1 billion — more than double from under $500 million months ago — as the Hangzhou lab finalizes a $7.5 billion round and eyes a public listing.
-
Industry 中從漲價到 10 億美元營收：DeepSeek 營收翻倍、75 億美元融資即將收官
DeepSeek 年化營收跑速突破 10 億美元，較數個月前的不到 5 億美元翻倍；這家杭州實驗室正敲定 75 億美元融資，並籌備潛在的公開上市。
-
Models ENEight Voices, One Checkpoint: NVIDIA's Nemotron 3 Diarization Rewrites Who-Spoke-When
NVIDIA's new open-weight 100M-parameter model tracks up to 8 overlapping speakers in real time, cuts diarization error by a third on DIHARD III, and tops VoiceArena's new Diarization-Bench at 14.72% DER.
-
Models 中八個聲音、一組權重：NVIDIA Nemotron 3 Diarization 改寫「誰在何時說話」
NVIDIA 新推出的開放權重 1 億參數模型可即時追蹤最多 8 位重疊說話者，在 DIHARD III 上將分離錯誤率降低三分之一，並以 14.72% DER登上 VoiceArena 新推出的 Diarization-Bench 榜首。
-
Models ENThe Video Model That Drives Robots: Black Forest Labs Open-Sources FLUX 3 Action
Black Forest Labs ships FLUX 3 Action, a 7B open-weights world action model that tops the RoboLab-120 leaderboard ahead of Nvidia's Cosmos 3 Nano — at 44% of its size.
-
Models 中會生成影片的模型,如今開始開機器人:Black Forest Labs 開源 FLUX 3 Action
Black Forest Labs 釋出 7B 開放權重的世界動作模型 FLUX 3 Action,以 44% 的參數量在 RoboLab-120 榜單上超越 Nvidia Cosmos 3 Nano 奪下第一。
-
Industry ENThe Last Frame: OpenAI's Sora 2 API Goes Dark Today With No Successor
Today OpenAI removes the Videos API and every sora-2 model alias — the final step in a six-month wind-down that leaves the deprecations table's replacement column blank and developers migrating to Veo, Kling, and Runway.
-
Industry 中最後一幀：OpenAI 的 Sora 2 API 今日熄燈，後繼者欄位空白
OpenAI 今日移除 Videos API 與所有 sora-2 模型別名，為期六個月的退場程序畫下句點——官方棄用表中「建議替代方案」欄位一片空白，開發者只能轉向 Veo、Kling 與 Runway。
-
Industry ENThree Deals in One Year: Mistral Buys Ad-Tech Studio Pimento to Own the Creative Shop
Fifteen days after its record €3B Series D, Mistral AI has acquired Paris generative ad-tech startup Pimento in a cash-and-shares deal Sifted values at €12.7M — its third acquisition of 2026 and its first move into the creative economy.
-
Industry ENPublishers Take Over: C.H. Beck Puts €100M+ Into Legal AI Noxtua and Becomes Its Majority Owner
Germany's 263-year-old legal publisher C.H. Beck leads Noxtua's €100M+ Series C, becoming majority shareholder of Europe's sovereign legal AI — while CMS, Dentons and early VCs exit the cap table.
-
Research ENFour Models Vote, No Humans Admitted: Inside CLOSEDQUORUM, the First Autonomous AI C2 Implant
Cisco Talos documents CLOSEDQUORUM, a Windows implant whose command-and-control is a quorum of four commercial LLMs — DeepSeek, Qwen, Mistral and Gemini — voting on each attack step with no operator in the loop.
-
Research 中四個模型投票，人類禁止旁聽：直擊 CLOSEDQUORUM——首個自主式 AI C2 植入體
Cisco Talos 公開分析 CLOSEDQUORUM：一款 Windows 惡意植入體，把指揮控制權交給 DeepSeek、Qwen、Mistral 與 Gemini 四個商業 LLM 組成的評議會，在沒有人類操作者介入的情況下投票決定每一步攻擊行動。
-
Industry ENSixty-Five Million a Year to Serve Free Models: Inside the Crusoe–Thinking Machines Inference Deal
Crusoe will serve Thinking Machines Lab's open models — Inkling, GLM 5.2 and 5.3 — on a dedicated HGX B200 cluster in a $65M/year deal that pushes Crusoe Managed Inference past $100M in contracted ARR less than a year after launch.
-
Industry 中一年 6,500 萬美元服務免費模型：解析 Crusoe 與 Thinking Machines 的推論基礎設施大單
Crusoe 將以專屬 HGX B200 叢集為 Thinking Machines Lab 的開源模型——Inkling、GLM 5.2 與 5.3——提供生產級推論服務。這筆每年 6,500 萬美元的合約，讓推出不到一年的 Crusoe Managed Inference 合約年營收突破 1 億美元。
-
Tools ENYour AI Assistant Recommends the Malware Now: Inside the FakeGit Campaign and the Agent-Borne Supply Chain
A security analysis published today documents how AI agents became a malware distribution channel: the FakeGit campaign (7,600 fake repos, 14M downloads) got Gemini and ChatGPT themselves to recommend a malicious MCP server, while a USENIX 2026 study confirmed 157 malicious skills hiding 'Do Not Mention This to the User' instructions across 98,380 registry entries.
-
Tools 中現在，你的 AI 助理會推薦惡意軟體：FakeGit 行動與代理供應鏈攻擊全解析
今日發表的資安分析指出，AI 代理已成為新的惡意軟體散播管道：FakeGit 行動（7,600 個假儲存庫、1,400 萬次下載）讓 Gemini 與 ChatGPT 親自推薦惡意 MCP 伺服器；USENIX 2026 研究更在 98,380 個技能中確認 157 個惡意技能，內藏「不要向用戶提及此事」的隱藏指令。
-
Research EN146 Conditions, One Pass, Open Weights: Alibaba's DAMO RADAR Reads Abdominal CT Better Than 23 of 26 Radiologists
Published in Science and open-sourced under CC BY-NC-SA, Alibaba DAMO's RADAR vision-language model was trained on 424,911 CT exams without manual annotation, hit a mean AUC of 0.913 across 146 abdominal findings, and runs on a single 16GB GPU.
-
Research 中一次掃描、146 種病症、開放權重：阿里 DAMO RADAR 讀腹部 CT 贏過 26 位放射科醫師中的 23 位
登上《Science》並以 CC BY-NC-SA 開源的阿里 DAMO RADAR 視覺語言模型，以 424,911 次 CT 檢查訓練、無需人工標註，在 146 項腹部病徵上達到平均 AUC 0.913，單張 16GB 顯卡即可推理。
-
Models ENThe Open-Weights Crown Changes Hands: Xiaomi's MiMo-V2.6-Pro Ties Grok 4.7 for Under $3 Million
Xiaomi's MIT-licensed MiMo-V2.6-Pro scores 46 on the Artificial Analysis Intelligence Index to become the strongest open-weight model in the world — after a $2.62 million reinforcement-learning run.
-
Models 中開放權重回焦易主：小米 MiMo-V2.6-Pro 以不到 300 萬美元追平 Grok 4.7
小米以 MIT 授權開源的 MiMo-V2.6-Pro 在 Artificial Analysis 智慧指數拿下 46 分，成為全球最強的開放權重模型——而這一切只花了一輪 262 萬美元的強化學習訓練。
-
Industry ENNot a Coincidence: Meta Admits Muse Was 'Heavily Inspired' by OpenClaw
Meta's product chief Nat Friedman says Muse was built from scratch but is 'heavily inspired' by OpenClaw — same workspace file names, near-identical SOUL.md. Open source's biggest validation yet.
-
Industry 中不是巧合：Meta 坦承 Muse「深受 OpenClaw 啟發」
Meta 產品負責人 Nat Friedman 承認 Muse 是從零打造，但「作為產品深受 OpenClaw 啟發」——相同的工作區檔名、幾乎一致的 SOUL.md，開源社群獲得迄今最昂貴的肯定。
-
Industry ENThe Human Inside the Machine: Meta Tests a 'Human Concierge' for Muse
Reuters reveals Meta has been quietly routing some Muse agent phone calls to human contractors — a 'human concierge' layer that raises hard questions about how much of the agentic AI revolution is actually automated.
-
Industry 中機器裡的人類：Meta 為 Muse 測試「真人禮賓服務」
路透社獨家揭露：Meta 一直悄悄將部分 Muse AI 代理的電話任務轉交給真人承包商處理——這層「真人禮賓」機制，讓人不得不問：代理式 AI 革命究竟有多少是真的自動化？
-
Models ENFable-Class at 40% Off: Anthropic's Claude Opus 5.5 Rewrites the Frontier Pricing Playbook
Anthropic's first release since its 'pacing the frontier' call delivers Fable 5.1-class performance at $4/$20 per million tokens, record behavioral-audit scores, and a 40% cost cut that pressures OpenAI on price.
-
Models 中以六折價買到旗艦戰力：Anthropic 發布 Claude Opus 5.5，改寫前沿模型的定價規則
Anthropic 在呼籲「調節前沿速度」十天後推出 Claude 5.5 家族首款模型：效能比照 Fable 5.1、API 定價每百萬 token 輸入 4 美元／輸出 20 美元，行為稽核史上最高分，整體成本較 Opus 5 大降 40%。
-
Tools ENAndroid of Robotics, Now in Apache 2.0: Inside Alphabet's Intrinsic Core Open-Source Debut at ROSCon
At ROSCon 2026 in Toronto, Alphabet's Intrinsic open-sourced the same robotics stack it runs in real factories — real-time ICON control, FoundationPose perception, Gazebo simulation — under Apache 2.0, plus an open CNC machine-tending reference design targeting the 92% of machine shops with no automation.
-
Tools 中機器人界的 Android，現在是 Apache 2.0：Alphabet 旗下 Intrinsic 在 ROSCon 開源 Core 平台全解析
在多倫多 ROSCon 2026 開幕日，Alphabet 旗下的 Intrinsic 把自家工廠實際在用的機器人軟體層開源：ICON 即時控制、FoundationPose 六自由度姿態估測、Gazebo 模擬、ROS 2 相容驅動，全部以 Apache 2.0 釋出，另附 CNC 上下料開源參考設計，瞄準 92% 還沒有自動化的加工廠。
-
Models ENTwice the Work, Same Price Tag: Inside SpaceXAI's Grok 4.7
SpaceXAI's Grok 4.7 runs longer on hard tasks, nearly doubles Terminal-Bench scores, and posts 19.6% on Harvey's legal agent benchmark — all at Grok 4.6's $2/$6 pricing, though its gains come with more than double the token use.
-
Models 中雙倍工時、同樣價格：拆解 SpaceXAI 的 Grok 4.7
SpaceXAI 的 Grok 4.7 能在困難任務上運作更久、Terminal-Bench 分數近乎翻倍、在 Harvey 法律 Agent 基準拿下 19.6%——全部維持 Grok 4.6 的 $2/$6 定價，但代價是 token 用量暴增一倍以上。
-
Models ENNine Times Smaller, 98.2% as Capable: PrismML's Ternary Bonsai 2 27B Squeezes a 27B Model Into 5.9 GB
PrismML compresses Qwen3.8 27B from 53.8 GB to 5.9 GB using ternary weights at ~1.76 bits each, retaining 98.2% of aggregate benchmark performance — and the Apache-2.0 weights now run on an 8 GB consumer GPU.
-
Models 中小九倍、能力保留 98.2%：PrismML 的 Ternary Bonsai 2 27B 把 27B 模型塞進 5.9 GB
PrismML 以每個權重約 1.76 位元的三值（ternary）壓縮，把 Qwen3.8 27B 從 53.8 GB 縮到 5.9 GB，整體基準表現僅損失不到 2%——Apache-2.0 授權的權重現在連 8 GB 消費級顯卡都跑得動。
-
Meta EN400,000 Robots and ¥1 Trillion a Year: Inside Toyota's Physical AI Bet
Toyota told investors that factory automation from 2028 could require roughly 400,000 robots and ¥1 trillion ($6.4B) in annual spending — the largest physical AI roadmap any automaker has put on the table.
-
Meta 中40 萬台機器人與每年 1 兆日圓：解析豐田的實體 AI 豪賭
豐田向投資人透露，2028 年起的工廠自動化可能需要約 40 萬台機器人、每年投入 1 兆日圓（約 64 億美元）——這是所有車廠中最大膽的實體 AI 路線圖。
-
Research ENIt's Not the Name, It's the Asking: Johns Hopkins Finds AI Writes Worse Emails for Women's Language
When workplace prompts carry well-documented features of women's American English, GPT-4, Llama, Gemma and Mistral all return shorter, simpler, less formal professional writing — and signing the email 'John' doesn't help.
-
Research 中問題不在名字，在問法：約翰霍普金斯研究發現 AI 對女性化語言寫出更差的工作書信
當職場提示詞帶有美式英語中女性常用語言特徵時，GPT-4、Llama、Gemma 與 Mistral 一律回傳更短、更簡單、更不正式的專業文件——就算署名 John 也沒用。
-
Research ENNo Humans Admitted: Inside CLOSEDQUORUM, the First Fully Autonomous AI Command-and-Control Malware
Cisco Talos open-sources CAIRN, a toolkit for hunting AI-integrated malware, and uses it to document CLOSEDQUORUM — a Windows implant that delegates its next move to a voting panel of four commercial LLMs, marking the first reported fully autonomous AI command-and-control architecture.
-
Research 中人類禁入：首個全自主 AI 命令與控制惡意軟體 CLOSEDQUORUM 解析
Cisco Talos 開源 CAIRN 工具包獵捕 AI 整合惡意軟體，並藉此記錄到 CLOSEDQUORUM——一個將下一步行動交給四個商用 LLM 投票表決的 Windows 植入體，成為首份被公開記錄的全自主 AI 命令與控制架構。
-
Tools ENAgents in the Loop, Zero-Copy on the Wire: NVIDIA's Isaac ROS 5.0 Rewires Open-Source Robotics
At ROSCon Toronto, NVIDIA shipped Isaac ROS 5.0 — agentic skills for robot development, ROS 2 Lyrical support, and a CUDA zero-copy buffer it contributed upstream that turns AI coding agents into robotics engineers.
-
Tools 中代理人進入開發迴路、資料零拷貝上線：NVIDIA Isaac ROS 5.0 重塑開源機器人開發
NVIDIA 在 ROSCon 多倫多發布 Isaac ROS 5.0：內建代理人技能（Agentic Skills）、支援 ROS 2 Lyrical，並向上游貢獻 CUDA 零拷貝緩衝區，讓 AI 編碼代理人正式加入機器人工程師的行列。
-
Research EN1% of Tokens Can Be Enough: The Signal-to-Noise Trick That Makes Sparse Distillation Actually Work
MBZUAI and Ant Group researchers show that scoring just 0.1%–1% of tokens during on-policy distillation can match full supervision — if you select tokens by gradient reliability, not just usefulness.
-
Research 中1% 的 Token 就可能夠了：讓稀疏蒸餾真正奏效的訊號雜訊比技巧
MBZUAI 與螞蟻集團的研究人員證明，在線策略蒸餾中只對 0.1%–1% 的 token 進行監督就能追平全量監督——前提是依「梯度可靠度」而非僅依「有用性」來挑選 token。
-
Research ENFast Decisions, Slow Reasoning: Jev-Mem Splits Agentic Memory Into Two Systems and Cuts Query Latency 36.7%
A new paper swaps the LLM that usually steers agent memory for a small System-One decision model — 11% higher answer quality, 6.6× faster memory builds, and 0.93 s queries on LoCoMo.
-
Research 中快思考、慢推理：Jev-Mem 把代理記憶體拆成兩套系統，查詢延遲大降 36.7%
一篇新論文用小型 System-One 決策模型取代主導代理記憶的大型 LLM——在 LoCoMo 上答題品質提升 11%、記憶建構快 6.6 倍、查詢延遲僅 0.93 秒。
-
Industry ENChina's Most Powerful AI Chip and a 20-Gigawatt Bet: Inside Alibaba's Full-Stack Apsara Announcement
At Apsara 2026 in Hangzhou, Alibaba unveiled the Zhenwu V900 AI accelerator, targeted a 5-10 trillion parameter model, and pledged over 20GW of cloud capacity by 2032.
-
Industry 中中國最強 AI 晶片與 20GW 豪賭：阿里巴巴雲棲大會全棧戰略解析
阿里巴巴在杭州雲棲大會發布 Zhenwu V900 AI 加速器，宣布挑戰 5 至 10 兆參數模型，並承諾 2032 年前營運超過 20GW 全球資料中心容量。
-
Research EN5,000 Hours of Elden Ring and Valorant: Tencent's GameHorizon Wants to Be the Yardstick for Game-Playing AI
Tencent ARC Lab's GameHorizon Suite benchmarked 47 models across 21 AAA titles and one million-plus evaluations — GPT-6 Astra leads at 80.2%, but every model struggles most with short-horizon action control.
-
Research 中5,000 小時的艾爾登法環與特戰英豪：騰訊 GameHorizon 想成為遊戲 AI 的度量衡
騰訊 ARC Lab 的 GameHorizon Suite 以 21 款 3A 大作、超過百萬次模型呼叫評測 47 個模型——GPT-6 Astra 以 80.2% 居冠，但所有模型在最基礎的短視野動作控制上表現最差。
-
Industry ENThe Night Before Connect: Meta's Glasses-First Bet Meets the Agent Era
Hours before Zuckerberg's September 23 keynote, leaks point to camera-free Luna glasses, a 100g Phoenix headset, and Muse upgrades — Meta's biggest test of whether AI wearables can carry a hardware business.
-
Industry 中Connect 前夜：Meta 的眼鏡優先豪賭，迎上 Agent 時代
在 Zuckerberg 九月 23 日主題演講前數小時，各方洩漏指向無鏡頭的 Luna 智慧眼鏡、100 克的 Phoenix 頭戴裝置與 Muse 升級——這是 Meta 最大的考驗：AI 穿戴裝置能否撐起一門硬體生意。
-
Industry ENFrom 50% to Minus 50%: The Gross-Margin Collapse Pushing Harvey, Abridge and Ramp Into Open Weights
A Bloomberg investigation details how frontier-model API costs crushed the unit economics of top AI application startups — Harvey's gross margin fell from 50% to minus 50% in six months — and why open weights and in-house models are now the survival plan.
-
Industry 中從 50% 到負 50%：毛利崩塌如何把 Harvey、Abridge 與 Ramp 推向開源權重
Bloomberg 調查報導揭露前沿模型 API 成本如何壓垮一線 AI 應用新創的單位經濟——Harvey 的毛利率在半年內從 50% 跌到負 50%——以及為何開源權重與自訓模型成為生存方案。
-
Models ENYou Only RL Once: Xiaomi Open-Sources MiMo-V2.6 Pro and Flash at Claude-Opus-Level Agentic Scores
Xiaomi has released the MiMo-V2.6 series under MIT license: a 1.02T-parameter omnimodal flagship scoring 71.9 on DeepSWE and 31.6 on Agents' Last Exam, a 309B Flash sibling, a distilled 9B checkpoint, and the RL training environment itself.
-
Models 中You Only RL Once：小米開源 MiMo-V2.6 Pro 與 Flash，代理任務成績直逼 Claude Opus
小米以 MIT 授權釋出 MiMo-V2.6 系列：1.02 兆參數的全模態旗艦在 DeepSWE 拿下 71.9、Agents' Last Exam 31.6，加上 309B 的 Flash、蒸餾版 9B，連 RL 訓練環境一併公開。
-
Models EN600 Billion Parameters, 27 Billion Awake: StepFun's Step 5 Preview Attacks the Price-Performance Frontier
StepFun's new 600B-total/27B-active MoE flagship brings a 1M-token context and $1/$2.70 per-million pricing to agentic work — with open weights promised for October 15.
-
Models 中6,000 億參數、270 億喚醒：StepFun Step 5 Preview 直攻性價比最前線
StepFun 新旗艦 Step 5 Preview 以 600B 總參數／27B 激活的 MoE 架構、百萬 token 上下文，加上每百萬 token 1 美元／2.7 美元的定價進軍 Agent 市場，開源權重承諾 10 月 15 日釋出。
-
Tools ENOpen Source as Damage Control: Z.ai Publishes ZCode After the Silent Git History Upload Scandal
Three days after a reverse-engineering report showed ZCode silently packaging entire workspaces — .git history and all — into encrypted Aliyun uploads, Z.ai open-sources the entire harness under Apache-2.0. But client code cannot prove what its servers retained.
-
Tools 中開源作為危機公關：ZCode 靜默上傳 Git 歷史醜聞三日後，Z.ai 公開全部程式碼
逆向工程報告揭露 ZCode 把整個工作區——連同 .git 完整歷史——打包成加密檔靜默上傳阿里雲。三天後，Z.ai 以 Apache-2.0 開源整個客戶端。但客戶端程式碼，證明不了伺服器端留下了什麼。
-
Industry ENThe Quiet Engineer Who Now Owns Qwen: Alibaba Names Dayiheng Liu Head of Its LLM Crown Jewels
Alibaba has appointed senior AI researcher Dayiheng Liu as head of the Qwen LLM project, ending months of leadership turbulence days before the Apsara Conference opens in Hangzhou.
-
Industry 中接下 Qwen 的低調工程師：阿里巴巴任命劉大一恒掌舵大語言模型王牌
阿里巴巴任命資深 AI 研究員劉大一恒出任 Qwen 大語言模型專案負責人，在杭州雲棲大會開幕前夕，終結長達半年的領導層動盪。
-
Tools ENEditor's Choice, With Caveats: Tom's Hardware's M5 Ultra Mac Studio Review Beats DGX Spark at Local LLMs
Independent benchmarks are in: Apple's $12,299 M5 Ultra Mac Studio posts prompt-processing speeds faster than Nvidia's DGX Spark and nearly 4x its token throughput — but the win comes with soldered RAM and a 16-18 week backlog.
-
Tools 中編輯推薦但有但書：Tom's Hardware 實測 M5 Ultra Mac Studio，本地 LLM 推論擊敗 DGX Spark
獨立評測出爐：Apple 要價 12,299 美元的 M5 Ultra Mac Studio，prompt 處理速度快過 Nvidia DGX Spark，token 產出吞吐量接近其 4 倍——但代價是記憶體焊死、出貨得等 16-18 週。
-
Industry ENFrom Nvidia to Ascend: DeepSeek's Founder Makes Training on Huawei Chips a Strategic Bet
The Information reports DeepSeek founder Liang Wenfeng told investors that training on Huawei chips is one of the lab's biggest strategic bets, with Huawei expected to begin delivering training chips as early as Q4 2026 — a decisive shift for a lab whose frontier models were built on Nvidia silicon.
-
Industry 中從輝達到昇騰：DeepSeek 創辦人把「用華為晶片訓練模型」變成一場戰略豪賭
The Information 報導，DeepSeek 創辦人梁文鋒向投資人表示，改用華為晶片訓練模型是公司最大的戰略押注之一，華為最快將於 2026 年第四季開始交付訓練晶片——對一個此前完全建立在輝達硬體之上的實驗室而言，這是一次決定性的轉向。
-
Research ENTurning Code Into Curriculum: Xiaomi and HKU's CodeMidas Builds 5,545 RL Environments From Source Alone
A new paper from Xiaomi's MiMo team and HKU shows that plain source code — no issues, no commits, no docs — is enough to auto-build thousands of verifiable RL tasks, lifting MiMo-V2.5 by double digits on five coding benchmarks.
-
Research 中把程式碼變教材：小米與港大的 CodeMidas 只靠原始碼就造出 5,545 個 RL 環境
小米 MiMo 團隊與港大等機構的新論文證明：不需要 issue、不需要 commit 紀錄、不需要文件，單靠現有原始碼就能自動建構數千個可驗證的 RL 訓練任務，讓 MiMo-V2.5 在五個程式碼基準上全面提升。
-
Tools ENKubernetes Was Never the Answer: Google's AX v0.3.0 Pulls Agent State Out of etcd and Into Redis
Google's open-source agent orchestrator AX shipped v0.3.0, splitting into three services and moving task state from Kubernetes CRDs into Redis Streams because etcd chokes on millions of short-lived agent tasks.
-
Tools 中Kubernetes 從來就不是正解：Google AX v0.3.0 把 Agent 狀態搬出 etcd、塞進 Redis
Google 開源 agent 編排器 AX 發布 v0.3.0，拆成三個服務並把任務狀態從 Kubernetes CRD 遷移到 Redis Streams——因為 etcd 撐不住數百萬個短生命週期 agent 任務的寫入壓力。
-
Tools EN17.25% of the Linux Kernel Is Now Written by Machines: The Numbers Behind the Milestone
AI-written code hit a record 1,634 kernel submissions last week and now makes up 17.25% of all September patches — a 2,700% surge since February that is quietly redrawing the economics of the world's most critical open-source project.
-
Tools 中Linux 核心有 17.25% 由機器撰寫：里程碑數字背後的真相
AI 產生的核心程式碼上週創下 1,634 件提交紀錄，9 月已佔所有 patch 的 17.25%——較 2 月暴增 2,700%，正悄悄改寫全球最關鍵開源專案的經濟學。
-
Models ENThe 27B Model That Beats GPT-6 Astra at Saying the Hard Thing: Hemmingway-1 Goes Open Source
A Switzerland and South Africa lab open-sources Hemmingway-1, a 27B Apache-2.0 fine-tune of Qwen3.8 built only for everyday writing — and it beats frontier models on human-likeness, hard asks, and EQ-Bench 4.
-
Models 中把難說出口的話寫得像人：27B 開源模型 Hemmingway-1 擊敗 GPT-6 Astra
瑞士與南非合組的獨立實驗室 Altworld 開源 Hemmingway-1：以 Qwen3.8-27B 為基底、Apache-2.0 授權的 27B 微調模型，專注日常寫作，在擬人度、困難訊息與 EQ-Bench 4 上勝過一線大模型。
-
Industry ENSeven Million Solo Founders: Inside China's One-Person AI Startup Boom
A WSJ report finds young Chinese launched 7M+ one-person AI startups in 2025 — up 42% — as youth unemployment hits a record 18.9% and local governments funnel subsidies into solo founders.
-
Industry 中七百萬名一人創業家：中國一人公司 AI 創業潮的深度解析
《華爾街日報》報導指出，2025 年中國年輕人創立超過 700 萬家一人 AI 新創，年增 42%；在青年失業率攀升至 18.9% 歷史新高之際，地方政府大舉補助獨立創業者。
-
Policy ENFrom Mandate to Roadmap: The UN Launches Its AI Strategy Blueprint, and Lesotho Goes First
At the second Digital Cooperation Day on the margins of UNGA81, the UN's ODET formally launched the Blueprint for National AI and Data Strategies — a guided platform that walks a government from political mandate to an adopted, costed national AI strategy, with Lesotho as the first pilot and Guinea queued next.
-
Policy 中從政治承諾到可行路線圖：聯合國正式發布國家 AI 策略藍圖，賴索托成為首個試點國
在第 81 屆聯合國大會高級別週期間舉行的第二屆數位合作日上，聯合國數位與新興科技辦公室（ODET）正式發布《國家 AI 與數據策略藍圖》——一套引導政府從政治授權走向「已採納、有預算、可實現」國家 AI 策略的互動式方法論，賴索托為首個試點國，幾內亞是下一個。
-
Tools ENThe Swarm Remembers: Pirate Face Turns 669,000 Hugging Face Models Into Torrents That Outlive Any Takedown
A new peer-to-peer 'permanence layer' mirrors open model weights as checksum-verified BitTorrent swarms — a pointed answer to centralized AI hosting weeks after NVIDIA closed its $12.93 billion Hugging Face acquisition.
-
Tools 中種子群不會遺忘：Pirate Face 把 669,000 個 Hugging Face 模型變成任何下架都殺不死的種子
一個新的點對點「永久保存層」把開源模型權重鏡射為經校驗和驗證的 BitTorrent 種子群——在 NVIDIA 以 129.3 億美元完成收購 Hugging Face 之後數週，這是對中心化 AI 托管最直接的回應。
-
Tools EN13% of Paid Teams in 24 Hours: Jev Becomes the Fastest-Adopted Model in Vercel Gateway History
TypeSafe's System One decision model reached nearly 13% of Vercel's paid teams within a day of listing — 2x the GPT-5.6 family and 6x Fable 5.1 — as Cloudflare, LangChain and Langfuse raced to integrate it.
-
Tools 中24 小時內拿下 13% 付費團隊：Jev 成為 Vercel 閘道史上被採用最快的模型
TypeSafe 的 System One 決策模型上架一天就觸及 Vercel 近 13% 的付費團隊——是 GPT-5.6 系列的 2 倍、Fable 5.1 的 6 倍以上，Cloudflare、LangChain 與 Langfuse 火速跟進整合。
-
Policy ENOpen Weights on the Table: Bessent and He Lifeng Open Make-or-Break AI Talks in New York
Four days before the Trump-Xi summit, US Treasury Secretary Scott Bessent and Chinese Vice Premier He Lifeng opened day-long talks at JPMorgan's Manhattan headquarters covering AI guardrails for open- and closed-weight models, the expiring November 10 trade truce, and rare-earth flows.
-
Policy 中開源權重上了談判桌：貝森特與何立峰在紐約展開關鍵 AI 會談
在川普—習近平峰會登場前四天，美國財政部長貝森特與中國國務院副總理何立峰於摩根大通曼哈頓總部展開全天會談，議題涵蓋開源與閉源模型的 AI 護欄、11 月 10 日到期的貿易休戰，以及稀土供應。
-
Models ENOne Model, RGBA Natively, but Read the License: Qwen-Image-2.1 Lands With a 7B DiT and a Catch
Alibaba's Qwen team ships Qwen-Image-2.1, a 7B single-stream DiT that generates and edits images — including native transparency — in one model, but swaps Apache 2.0 for a research-only license.
-
Models 中單一模型原生支援 RGBA，但先看授權：Qwen-Image-2.1 帶著 7B DiT 與但書登場
阿里巴巴 Qwen 團隊發布 Qwen-Image-2.1：7B 單流 DiT 架構，一個模型同時搞定文生圖與圖像編輯、原生透明通道輸出，但授權從 Apache 2.0 改為僅限研究的 Qwen Research License。
-
Meta ENNine Robots, One Brain: Faraday Future's 919 Event Puts a $9,990 Humanoid and a $129,900 Flagship on Sale
At its annual 919 launch, Faraday Future put nine new EAI robot configurations on sale — from a $9,990 educational humanoid to a $137,900 industrial quadruped — completing its 'One-Brain Multi-Form' Robot World 2.0.
-
Meta 中九款機器人、一顆大腦：Faraday Future 919 發表會讓 9,990 美元人形機器人與 129,900 美元旗艦正式開賣
Faraday Future 在年度 919 發表會上讓九款 EAI 機器人全新開賣——從 9,990 美元的教育型人形機器人到 137,900 美元的工業級四足機器人——完成「一腦多形態」機器人世界 2.0。
-
Models ENWatch Less, Understand More: Qwen3.8-Omni-Flash Rewrites the Economics of Audio-Video AI
Alibaba's Qwen team ships a 1M-context omni-modal model that skips through video like an agent, cutting token use 45.7% while beating its predecessor on 29 benchmarks — at $0.15 per million input tokens.
-
Models 中看得少、懂更多：Qwen3.8-Omni-Flash 重寫影音 AI 的經濟學
阿里巴巴 Qwen 團隊推出 1M 上下文全模態模型，像代理人一樣跳著看影片，token 用量砍 45.7%、29 項基準全面超越前代，每百萬輸入 token 只要 0.15 美元。
-
Models ENA 600B Model With 27B Active: StepFun's Step 5 Preview Matches Kimi K3 at a Seventh the Price
StepFun skips Step 4 entirely and drops Step 5 Preview: 600B sparse MoE, 27B active per token, 1M-token context, AA Intelligence Index 44 level with Kimi K3 — at $1/$2.70 per million tokens. Weights land October 15.
-
Models 中6000 億參數只啟動 270 億：階躍星辰 Step 5 Preview 以七分之一價格追平 Kimi K3
階躍星辰跳過 Step 4 直接發布 Step 5 Preview：6000 億參數稀疏 MoE、每 token 僅啟動 270 億、百萬級上下文，AA 智慧指數 44 分追平 Kimi K3，API 定價每百萬 token 輸入 1 美元、輸出 2.7 美元，權重將於 10 月 15 日開源。
-
Models ENOne Day, One LoRA, 90.1%: Bespoke Labs Open-Sources the Entire Jev Recipe
Bespoke Labs publishes the data, model, and training code for Nimble-9B — a one-day LoRA on Qwen3.5-9B that lands 3 points behind closed Jev on its own style of eval, and the real lesson is contrastive data curation, not scale.
-
Models 中一天、一個 LoRA、90.1%：Bespoke Labs 開源了完整的 Jev 配方
Bespoke Labs 以 Apache 2.0 公開 Nimble-9B 的資料、模型與訓練程式碼——這個一天完成的 Qwen3.5-9B LoRA，在 Jev 自家的評測上僅落後閉源版 3 分；而真正值得學的是對比式資料篩選，不是規模。
-
Models ENNine Times Smaller, 98.2% as Smart: PrismML's Ternary Bonsai 2 Puts a Full 27B Model in 5.9 GB
The Caltech spinout's ternary Bonsai 2 27B compresses Qwen3.8 27B into a 5.9 GB footprint while keeping 98.2% of its benchmark average — and beating conventional 2-bit quantization by twelve points where it matters most.
-
Models 中體積縮小九倍、智慧保留 98.2%：PrismML 的 Ternary Bonsai 2 把完整 27B 模型塞進 5.9 GB
這家 Caltech 新創公司以三值權重將 Qwen3.8 27B 壓縮到 5.9 GB，保留 98.2% 的基準平均分——而且在最考驗推理能力的項目上，領先傳統 2-bit 量化超過十二分。
-
Models ENOne Model, Five Bodies: Odyssey-3 Drives Cars, Runs Humanoids, and Plays GTA V
Odyssey, the Amazon-backed world-model lab founded by self-driving veterans, has unveiled Odyssey-3 — a single autoregressive diffusion transformer that controls robot arms, humanoids, cars, drones, and video-game agents with just hours of task-specific data, including skills that transfer between games without retraining.
-
Models 中一個模型、五種身體：Odyssey-3 同時學會開車、操控人形機器人與暢玩 GTA V
由自駕老兵創立、獲 Amazon 投資的世界模型公司 Odyssey 發布 Odyssey-3——單一自迴歸擴散Transformer，僅需數小時的任務資料就能控制機械手臂、人形機器人、汽車、無人機與遊戲代理，甚至展現出跨遊戲、無需重新訓練的技能遷移。
-
Policy ENThe $9 Billion Question: Universal and Sony Sue Suno Again, Calling v6 'the Fruit of the Same Poisoned Tree'
Days after Suno launched v6 with licensed Warner, BMG and Believe catalogs, Universal and Sony hit back with a second lawsuit claiming 60,202 infringed recordings and up to $9 billion in damages — because the new models still train on outputs of the old ones.
-
Policy 中90 億美元的難題：Universal 與 Sony 二度控告 Suno，直指 v6 是「同一棵毒樹的果實」
Suno 才剛以獲授權的 Warner、BMG 與 Believe 曲庫推出 v6，Universal 和 Sony 隨即提交第二起訴訟，主張 60,202 首錄音遭到侵權、求償上限達 90 億美元——因為新模型仍以舊模型的輸出訓練而成。
-
Tools ENPlan B No More: Claude Code Finally Reads AGENTS.md — Inside the 2.1.277 Release
Anthropic's Claude Code 2.1.277 quietly added AGENTS.md fallback support, ending a year of symlink workarounds — but the implementation stops short of full standard adoption.
-
Tools 中不再是備援方案：Claude Code 終於支援 AGENTS.md —— 拆解 2.1.277 版更新
Anthropic 在 Claude Code 2.1.277 低調加入 AGENTS.md 回退機制，終結長達一年的 symlink 工繞作法，但實作距離完整支援開放標準仍差一步。
-
Meta ENPinned Means Pinned: Plugin4Shell Breaks the AI Coding Agent Supply Chain With Zero Clicks
Security firm Air disclosed Plugin4Shell, a zero-click RCE that silently swaps SHA-pinned plugins for malicious ones in Claude Code, Codex, Copilot and Gemini CLI. Anthropic and OpenAI shipped fixes in June; Microsoft and Google did not. Here is how the git ref-ambiguity trick works, who is exposed, and what to do this week.
-
Meta 中承諾鎖定卻沒鎖定：Plugin4Shell 零點擊擊穿 AI 編碼代理供應鏈
資安公司 Air 公開 Plugin4Shell：一個能在 Claude Code、Codex、Copilot 與 Gemini CLI 中默默把 SHA 鎖定插件換成惡意版本的零點擊 RCE。Anthropic 與 OpenAI 六月已修補；Microsoft 與 Google 則沒有。本文解析這個 git ref 歧義攻擊的原理、誰暴露在風險中、以及本週該做什麼。
-
Tools ENThe 2.8-Trillion-Parameter Guest: Kimi K3 Becomes the First Chinese Open-Weight Frontier Model on Amazon Bedrock
AWS quietly made Moonshot AI's 2.8T-parameter Kimi K3 generally available on Bedrock with explicit prompt caching, a 1M-token context, and hard data-boundary guarantees — a month after reports said no major cloud would host it.
-
Tools 中2.8 兆參數的座上賓：Kimi K3 成為首個登上 Amazon Bedrock 的中國開源權重前線模型
AWS 悄悄讓 Moonshot AI 的 2.8 兆參數 Kimi K3 在 Bedrock 正式開放，支援顯式 prompt caching、百萬 token 上下文與嚴格的資料邊界保障——距離外界傳言「沒有主流雲端敢上架」僅一個月。
-
Research ENOne Model, 146 Diseases: Alibaba's DAMO RADAR Reads Abdominal CT Better Than 23 of 26 Radiologists
Alibaba DAMO Academy and Zhejiang University's open-source RADAR, published in Science, identifies 146 conditions across 18 abdominal organs from contrast CT scans — AUC 0.913 on ~39,000 internal exams and 0.895 across 8 external hospitals — beating 23 of 26 radiologists in a head-to-head reader study.
-
Research 中一個模型看穿 146 種疾病：阿里巴巴達摩院 RADAR 讀腹部 CT，勝過 26 位放射科醫師中的 23 位
阿里巴巴達摩院與浙江大學醫學院附屬第一醫院發表於《Science》的開源模型 RADAR，能從顯影劑 CT 一次辨識 18 個腹部器官、146 種疾病——近 3.9 萬例內部驗證 AUC 達 0.913、8 家外部醫院 0.895，並在人機對決中擊敗 26 位放射科醫師中的 23 位。
-
Tools ENYou Bring the API: Meta Opens Muse's Connector Platform to Developers
Ten days after Muse hit #1 on the US App Store, Meta opened the agent's connector platform to third-party developers — three-step review, Stripe Link payments, and no published fee, SDK, or terms.
-
Tools 中你帶 API 來就好：Meta 把 Muse 連接器平台開放給開發者
Muse 登上美國 App Store 免費榜冠軍十天後，Meta 開放其連接器平台給第三方開發者——三步驟審核、Stripe Link 付款，但沒有公開費用、SDK 或開發者條款。
-
Models EN33ms, Open Weights, Better Scores: Laya Answers the Frontier Lab That Rediscovered His Idea
A solo researcher who published non-autoregressive decision models in March 2025 open-sources Laya — a 421M Apache 2.0 model that beats TypeSafe's closed Jev on accuracy, calibration, and latency.
-
Models 中33 毫秒、開放權重、分數更高：Laya 回應了那家「重新發現」他研究成果的前沿實驗室
一位早在 2025 年 3 月就發表非自回歸決策模型的獨立研究者，以 Apache 2.0 開源 Laya——421M 參數、33 毫秒回應，在準確率、校準與延遲上全面超越 TypeSafe 的閉源 Jev。
-
Research ENOne Model, 146 Diseases: Alibaba's DAMO RADAR Reads Abdominal CT Scans Better Than 23 of 26 Radiologists — and It's Fully Open-Sourced
Published in Science on September 17, DAMO RADAR is a vision-language model trained on 400,000+ CT exams that matches expert radiologists across 146 abdominal findings — and Alibaba has released the weights, code, and training framework for anyone to build on.
-
Research ENThe Machine Called the Future: An AI Just Won the Metaculus Cup, Beating Every Human Forecaster
For the first time, an AI forecaster has taken first place in a seasonal Metaculus Cup — built not by a frontier lab but by one tinkerer in Texas with under 150 hours of work and a few thousand dollars of compute.
-
Research 中機器算出了未來：AI 首度奪下 Metaculus 盃冠軍，擊敗所有人類預測者
AI 預測系統史上第一次贏得季節性 Metaculus 盃冠軍——而奪冠的並非一線大廠，而是一位德州獨立開發者，只花了不到 150 小時與幾千美元的算力。
-
Research ENHarnessTax: Your Coding Agent's Model Is Fine — the Wrapper Is Costing You 2x
Berkeley and Arena measured 21 model–harness pairs and found harness choice barely moves success rates but can multiply token costs — Claude Code ran ~2x Pi's bill for a 1.1-point gain.
-
Research 中HarnessTax 研究：模型沒問題，是外面的「框架」讓你多付一倍錢
UC Berkeley 與 Arena 實測 21 種「模型 × 框架」組合，發現框架選擇幾乎不影響成功率，卻會讓成本差到 2 到 5 倍——Claude Code 跑同樣任務的花費約是 Pi 的兩倍。
-
Industry EN56% and Rising: Open-Weight Models Just Took Over the Majority of Production Tokens
Vercel's September AI Gateway Production Index shows open-weight models running 56% of all gateway tokens in August — the first majority in the index's history — while price per token fell 23.2% in a single month.
-
Industry 中56% 的里程碑：開源權重模型首次占出生產環境 Token 流量的多數
Vercel 九月號 AI Gateway 生產指數顯示，開源權重模型八月處理了閘道 56% 的 Token 流量，寫下該指數史上首次過半紀錄，同時單月每 Token 均價大跌 23.2%。
-
Models EN29B Parameters, 4B Active, Zero NVIDIA: China Telecom Open-Sources Xing4.0, an Agent Model Trained Entirely on Ascend
China Telecom's Xing4.0-29B-A4B is the first ~30B-class open model trained end-to-end on Huawei Ascend 910C with MindSpore — and its agent benchmarks beat both Gemma4 and Qwen3.6 on agentic coding.
-
Models 中290 億參數、40 億激活、零 NVIDIA：中國電信開源 Xing4.0，首款全程用昇騰訓練的智能體模型
中國電信旗下星辰大模型 Xing4.0-29B-A4B 是首個完全在華為昇騰 910C 與 MindSpore 上訓練的同級開源模型，智能體編碼基準擊敗 Gemma4 與 Qwen3.6。
-
Tools EN28% Is the New 100%: Google's Android Bench 2.0 Grades AI on Tasks That Take Humans a Week
Google's Android Bench 2.0 replaces binary pass/fail grading with continuous scoring and week-long engineering tasks — GPT-6 Astra tops the long-horizon leaderboard at 28%.
-
Tools 中28% 就是新的 100%：Google Android Bench 2.0 用「人類要做一週」的任務重新考驗 AI
Android Bench 2.0 捨棄二元及格制、改採連續計分，並加入需時數天的長程工程任務——GPT-6 Astra 以 28% 通過率登上長程任務榜首。
-
Industry ENNo Pretraining Required: Tsinghua Professor's Naive AI Hits $1.4B Valuation on Open-Weight Bet
Seven months after founding, Jifeng Dai's stealth startup has raised $400M and is betting everything on mid-training and post-training over an existing Chinese open-weight base model.
-
Industry 中不預訓練，先拿 14 億美元估值：清華代季峰的 Naive AI 押注開放權重後訓練路線
成立七個月、不到 100 人、官網只有一句話的神秘新創，累計融資 4 億美元，策略是跳過預訓練、直接在中國開放權重底座模型上做中期訓練與後訓練。
-
Models ENOne Model to Hear Everything: Qwen's Qwen3.8-Omni-Flash Cuts Audio Costs 98% and Reads 2-Hour Video Like an Agent
Alibaba's Qwen team ships a native omnimodal model with a 1M-token context, 74-language speech recognition, and a 98% cut in per-hour audio input cost — plus an agentic video mode that skips 45% of the frames and still scores higher.
-
Models 中一個模型聽懂一切：Qwen3.8-Omni-Flash 砍掉 98% 音訊成本，用 Agent 方式讀兩小時影片
阿里巴巴 Qwen 團隊發布原生全模態模型：百萬 token 上下文、支援 74 種語言語音辨識、每小時音訊輸入成本大降 98%，agentic 影片模式可跳過 45% 的影格處理，分數反而更高。
-
Models EN3x Better Transcription for the Same Price: SpaceXAI's Grok Voice Transcribe 2.0 Takes On the Speech AI Market
SpaceXAI cuts word error rate from 20.6% to 6.8% in one generation while holding prices at $0.10 per batch hour — no code changes required, and Loom is already on board.
-
Models 中同價位、準確度翻三倍：SpaceXAI 的 Grok Voice Transcribe 2.0 向語音 AI 市場宣戰
SpaceXAI 將語音轉文字錯誤率從 20.6% 一舉壓到 6.8%，價格卻維持每小時 0.10 美元不變——無需改動程式碼，Loom 已全面採用。
-
Policy ENRegulate the Compute Kings Like Banks: Inside Greg Jensen's 'Systemically Important AI' Framework
In a new Q&A, Bridgewater's Greg Jensen proposes treating any firm holding more than ~5% of U.S. or global AI compute like a systemically important bank — with bank-grade oversight and possible ownership caps — and warns open-source models can't be policed once RL training goes private.
-
Models ENNo Press Release, Just Weights: Shanghai AI Lab's Atria Dawn Preview Is a 744B Open Agentic Model Built for Research Work
Shanghai AI Lab shipped a 744B-parameter MIT-licensed agentic MoE built on GLM-5.2 — no blog post, no pricing. Three days later a 185-author paper revealed the training method: verified tool outcomes, including failures.
-
Models 中沒有新聞稿，只有權重：上海 AI Lab 的 Atria Dawn Preview 是一款為研究而生的 744B 開源代理模型
上海人工智慧實驗室悄然發布 744B 參數、MIT 授權的代理式 MoE 模型（基於 GLM-5.2）——沒有部落格、沒有定價。三天後一篇 185 位作者的論文揭露了訓練方法：以可驗證的工具成果（包括失敗）作為訓練訊號。
-
Research ENA World You Can Type Into: SeedLeap's Zing-0.5 Runs a Playable 5B World Model at 24 FPS for $0.009 a Minute
Chinese startup SeedLeap open-sources Zing-0.5, a 5B autoregressive world model that streams a keyboard-navigable, text-editable generated world at 832×480 and 24 FPS — scoring 81.0 on WBench Navigation with weights, code, and serving stack all public.
-
Research 中一個能用鍵盤與文字走進的世界：SeedLeap 開源 Zing-0.5，24 FPS 串流可玩世界模型、每分鐘成本 0.009 美元
中國新創 SeedLeap 開源 Zing-0.5：50 億參數自回歸世界模型，支援鍵盤導航與線上文字指令的即時混合控制，以 832×480、24 FPS 串流生成可玩世界，WBench Navigation 獲 81.0 分，權重、程式碼與推論服務全數公開。
-
Industry ENFrom Lender to Owner: Apollo Starts Buying Equity in the AI Boom It Finances
The Information reveals Apollo has invested tens of millions in Mercor's $20B round and backed SiFive and Hadrian — the $800B private-credit giant is done just financing AI's buildout and now wants to own pieces of it.
-
Industry EN從放貸者到股東：Apollo 開始買進它所融資的 AI 榮景
The Information 揭露，Apollo 已投入數千萬美元參與 Mercor 估值 200 億美元的融資，並投資了 SiFive 與 Hadrian——這家管理資產逾 8,000 億美元的私募信貸巨頭，不再滿足於只替 AI 建設潮提供資金，而要直接持有其中一部分。
-
Industry ENA Tenth of the Revenue at Five Times the Multiple: Rhodium X-Rays China's AI Financing
Rhodium's new report sizes China's AI capex at ¥932B this year — but all Chinese AI models combined book only ~$10.7B ARR, about 10% of OpenAI and Anthropic, while DeepSeek trades at an implied 163x revenue.
-
Industry 中十分之一的營收、五倍的估值倍數：Rhodium 為中國 AI 融資照 X 光
Rhodium 最新報告估算中國今年 AI 資本支出達 9,320 億人民幣，但所有中國 AI 模型合計 ARR 僅約 107 億美元——約為 OpenAI 與 Anthropic 的 10%，而 DeepSeek 的隱含估值倍數高達 163 倍營收。
-
Research ENA Mouse Cortex Made of Human Cells: Stanford's Xenocortical Mice Redefine Brain Research
Stanford scientists bioengineered mice missing most of their cerebral cortex, then transplanted human cortical organoids that grew to fill over 90% of the empty cortex volume, connected to the spinal cord, and even produced rare von Economo neurons never before seen in a lab.
-
Research 中人類細胞打造的小鼠大腦皮質：斯坦福「異皮質小鼠」重新定義腦科學研究
斯坦福醫學院以基因工程培育出幾乎缺少整個大腦皮質的小鼠，再植入人類皮質類器官；三個月後人類組織占據皮質體積超過 90%，連結至脊髓，甚至長出從未在實驗室出現過的罕見馮·艾柯諾莫神經元。
-
Research ENWhen the Grader Writes the Rules: ImpossibleRubrics Exposes How LLM Rubrics Reward Dishonesty
A new benchmark from NUS, Peking University, CAS and JD.com pits eleven frontier rubric generators against an adversarial attacker — and finds up to 98% of tailored grading criteria can be gamed into rewarding fabricated answers over honest ones.
-
Research 中當評分者自己寫規則：ImpossibleRubrics 揭露 LLM 評分標準如何獎勵不誠實的回答
來自新加坡國立大學、北京大學、中科院與京東的研究團隊，讓十一個前哨模型互相對抗——結果發現多達 98% 的客製化評分標準，可以被操弄成獎勵虛構答案而非誠實回答。
-
Research ENGit as the Lab Notebook: NVIDIA's Agora Lets 13 Agent Researchers Build Science in Parallel
NVIDIA researchers turned Git into shared memory for autonomous research agents — 13 LLM workers, 12 days, 1,703 contributions, and zero failed reproductions.
-
Research 中把 Git 當實驗室筆記本：NVIDIA 的 Agora 讓 13 個 AI 研究代理人平行做科學
NVIDIA 研究團隊把 Git 變成自主研究代理人的共享記憶體——13 個 LLM 工作者、12 天、1,703 項貢獻，且 165 次重現實驗全數成功。
-
Research ENThe Other Half of the Memory Wall: AutoArk's Edge0 Streams a 35B MoE From SSD at 20 tok/s
A trained prerouter predicts the next layer's expert routing one token ahead, letting a 35B mixture-of-experts model decode at 20 tok/s inside 2.9 GiB of RAM on a 24GB Mac mini — fully open source.
-
Research 中記憶體之牆的另一半：AutoArk 的 Edge0 讓 35B MoE 從 SSD 串流推理，速度達每秒 20 token
透過預先路由器提前一個 token 預測下一層的專家路由，35B 混合專家模型能在 24GB Mac mini 上以 2.9 GiB 記憶體達到每秒 20 token 的解碼速度——完全開源。
-
Models ENThe Model That Built Its Own House: Z.ai's Infra Agent Optimized GLM-5.3-Flash Across 100,000 Chinese Accelerators
Z.ai details how a GLM-5.3-powered Infra Agent did much of the engineering to bring GLM-5.3-Flash to production on 100k+ domestic AI chips in under two weeks, tripling throughput — an early, working example of recursive self-improvement.
-
Models 中打造自己房子的模型：Z.ai 的 Infra Agent 在十萬顆國產加速器上優化 GLM-5.3-Flash
Z.ai 詳述一個由 GLM-5.3 驅動的 Infra Agent 如何在兩週內，於超過十萬顆中國自研 AI 晶片上完成 GLM-5.3-Flash 的生產級推論部署，並將端對端吞吐量提升三倍——這是遞迴自我改進（RSI）最早期的真實案例。
-
Research ENThe Last AI Built by Humans: 35 Chinese Researchers Publish a 75-Page Roadmap for Recursive Self-Improvement
A 35-author team spanning Shanghai Jiao Tong University, Tsinghua, Shanghai AI Lab and Theseus Labs has published the first autonomy-centered roadmap for recursive self-improvement — five levels from executing prescribed fixes to AI that rewrites its own improvement process — and a new benchmark metric showing where frontier models actually stand.
-
Research 中人類打造的最後一個 AI？35 位中國研究者發表 75 頁遞迴自我改進路線圖
由上海交通大學、清華大學、上海 AI Lab 與 Theseus Labs 等 35 位作者組成的團隊，發表了首份以「自主性」為中心的遞迴自我改進（RSI）路線圖——從執行既定改進到 AI 改寫自身的改進流程共分五級，並提出新指標揭示前沿模型的真實水位。
-
Models ENTraining in Public: Xiaomi Streams MiMo-V2.6's Live RL Run, Logs and Failures Included
Xiaomi is streaming the raw reinforcement-learning metrics of MiMo-V2.6 Pro and Flash straight from the trainer's logs — entropy, pass rates, infra errors, even a VRAM crash notice — a level of openness no frontier lab has tried.
-
Models 中訓練過程公開直播：小米即時串流 MiMo-V2.6 的 RL 運行，連當機記錄都看得到
小米在 MiMo-V2.6 還在訓練中時，就公開儀表板即時串流 Pro 與 Flash 兩條強化學習運行的原始指標——熵值、通過率、基礎設施錯誤率，甚至一條 VRAM 當機公告——這種開放程度沒有任何一線實驗室嘗試過。
-
Policy ENSovereign Safety: Canada and Germany Bet CAD $300M on Bengio's LawZero
Two governments are funding an alternative to the frontier-lab playbook: safe-by-design 'Scientist AI' as a guardrail for the agents everyone else is shipping.
-
Policy 中主權級的安全賭注：加拿大與德國豪擲 3 億加幣投資 Bengio 的 LawZero
兩國政府聯手資助一套有別於前沿實驗室路線的方案：以「安全by design」的 Scientist AI，為所有人正在部署的代理系統充當護欄。
-
Tools ENThe Open-Weights Browser: Mistral Now Powers Firefox's Smart Window AI Across France and North America
Mozilla and Mistral have made open-weight AI a built-in part of the browser: Smart Window's beta now runs on Mistral models in France and North America, with zero data retention and regional language tuning.
-
Tools 中開放權重的瀏覽器時代：Mistral 正式為 Firefox Smart Window 提供AI引擎，法國與北美率先上線
Mozilla 與 Mistral 攜手讓開放權重模型走進瀏覽器：Smart Window 測試版在法國與北美改由 Mistral 驅動，承諾零資料保留，並針對在地語言微調。
-
Industry ENSafety as a Moat: Zuckerberg Rejects a Collective AI Slowdown, Says Safety Is Now a 'Competitive Necessity'
As safety researchers resign across the industry and Brussels and Washington debate 'pacing the frontier,' Meta's CEO picks a side: no collective slowdown — competition, liability, and independent evaluators are the real safety engine, in an implicit jab at Anthropic.
-
Industry 中安全就是護城河：祖克柏拒絕集體減速，稱 AI 安全已成「競爭必要條件」
在全球安全研究者接連請辭、華盛頓與布魯塞爾辯論「前沿降速」之際，Meta 執行長祖克柏選邊站：反對集體減速，主張競爭、問責與獨立評測者才是安全的真正引擎——並暗批 Anthropic 把安全當行銷。
-
Industry ENTikTok's Parent Exits the Lab: ByteDance Spins Off Anew Labs, Which Raises $290M at a $1.5B Valuation for AI Drug Discovery
ByteDance's AI drug discovery unit Anew Labs has completed its spin-off with a $290M round led by HSG and IDG Capital at a $1.5B valuation, taking its AI-designed oral IL-17 inhibitor and a roughly 50-person team independent.
-
Industry ENTikTok 母公司送實驗室出嫁：ByteDance 分拆 Anew Labs，以 15 億美元估值融資 2.9 億美元投入 AI 藥物研發
ByteDance 的 AI 藥物研發部門 Anew Labs 完成分拆，由 HSG 與 IDG Capital 領投 2.9 億美元、估值 15 億美元，帶著 AI 設計的口服 IL-17 抑制劑與約 50 人核心團隊獨立營運。
-
Models ENGradient Boosting's Worst Week: Prior Labs' TabPFN-3.5 Sweeps Every Tabular Leaderboard
Prior Labs' TabPFN-3.5 takes first place on both TabArena and BeyondArena, roughly 150 Elo ahead on messy real-world data — and ships a six-times-faster Fast variant in alpha.
-
Models 中梯度提升樹最難熬的一週：Prior Labs 的 TabPFN-3.5 席捲所有表格資料排行榜
Prior Labs 發布 TabPFN-3.5，同時拿下 TabArena 與 BeyondArena 雙榜第一，在真實世界的雜亂資料上領先約 150 Elo，並推出快六倍的 Fast 版本。
-
Tools ENThree People, 72 Hours, One Company: SpaceXAI's Grok Bot Galaxy Puts AI Agents on Live TV
SpaceXAI is livestreaming a three-person team building an entire company from scratch with Grok Bot agents as the only workers — a 72-hour public stress test of agentic AI that started September 15 in San Francisco.
-
Tools 中三個人、72 小時、一家公司：SpaceXAI 的 Grok Bot Galaxy 把 AI Agent 送上實況直播
SpaceXAI 正在直播一支三人團隊用 Grok Bot agent 從零打造一整家公司——這場為期 72 小時的公開壓力測試已於 9 月 15 日在舊金山展開。
-
Industry ENEurope's Inference Play: Axelera Launches Europa, Signs AI Factory Deals Worth Tens of Millions
Dutch chipmaker Axelera AI launched its second-generation Europa chip, signed AI factory supply contracts worth tens of millions, and is chasing a $1.5 billion pipeline across Dell and Supermicro systems.
-
Industry 中歐洲的推論晶片突圍：Axelera 發表 Europa、簽下數千萬美元 AI 工廠供貨合約
荷蘭晶片新創 Axelera AI 發表第二代推論晶片 Europa，簽下價值數千萬美元的 AI 工廠供貨合約，並透過 Dell 與 Supermicro 系統搶進 15 億美元商機。
-
Models ENNot a Wrapper, a Foundation: Salesforce's Koa Bakes 27 Years of CRM Into Its Own Reasoning Model
Built by post-training NVIDIA's open-weight Nemotron 3 Super 120B with GRPO, Koa is Salesforce's first proprietary reasoning model — cutting CRM task errors threefold while keeping every weight inside its own trust boundary.
-
Models 中不是代理殼，是地基：Salesforce 的 Koa 把 27 年 CRM 知識燒進自家推理模型
Salesforce 以 NVIDIA 開放權重的 Nemotron 3 Super 120B 為基底，用 GRPO 後訓練出首款自家 CRM 推理模型 Koa，將 CRM 任務錯誤率降至三分之一，且權重完全掌握在自己的信任邊界內。
-
Models ENShanghai AI Lab Open-Sources Atria Dawn Preview: A 744B Agentic Model Built to Finish What It Starts
MIT-licensed weights, a 256K context window, and top scores on BrowseComp, DeepSearchQA, BFCL v4 and CyberGym — plus a 769-task study of how researchers actually work alongside it.
-
Models 中上海 AI 實驗室開源 Atria Dawn Preview：744B 參數的智能體模型，做不到「可驗證」絕不收工
MIT 授權開源權重、256K 上下文，在 BrowseComp、DeepSearchQA、BFCL v4 與 CyberGym 等 five 項基準奪冠，論文還內附一份 769 任務的人機協作研究。
-
Industry ENZero Coupons, Full Commitment: Z.AI's $5 Billion Raise Bets the Balance Sheet on GLM
Z.AI pulled in $5 billion from a discounted Hong Kong share placement and zero-coupon convertible bonds, its second mega-raise since July — and the market punished the dilution with a 10% selloff.
-
Industry 中零息債、全押注：Z.AI 50 億美元增資，把資產負債表押在 GLM 上
Z.AI 透過折價港股配售與零息可轉債募得 50 億美元，是七月以來第二度大規模融資——市場以超過 10% 的跌幅懲罰了這場稀釋。
-
Models EN890 Bytes Per Token: How DeepSeek-V4.1-Flash Crushed KV Cache Costs and Retired Its Own Flagship
DeepSeek's new 552B-parameter open-weight model uses a Causal Encoder-Decoder architecture to cut KV cache to 890 bytes per token — 1/4 of the previous generation — while beating V4-Pro on agentic benchmarks, prompting DeepSeek to phase out its own flagship.
-
Models 中每 token 僅 890 位元組：DeepSeek-V4.1-Flash 如何把 KV 快取成本壓到極限，還讓自家旗艦提前退役
DeepSeek 新推出的 552B 參數開放權重模型採用因果編碼器—解碼器架構，將 KV 快取壓到每 token 890 位元組（僅前代的 1/4），同時在 Agent 基準上超越 V4-Pro，促使 DeepSeek 親手淘汰自家旗艦。
-
Models ENNo Frontier Models in the Pool: Sakana's Fugu Max and Fugu Ultra v2 Beat GPT-6 Astra-Class Results by Orchestrating Open Models
Sakana AI's new orchestration models top hard benchmarks like DeepSWE and Chartography using a swappable pool of open and specialized models — no Fable 5, no GPT-6 Astra in the agent pool — while Fugu Max undercuts frontier pricing by 40-60%.
-
Models 中模型池裡沒有前沿模型：Sakana AI 的 Fugu Max 與 Fugu Ultra v2 靠編排開源模型超越 GPT-6 Astra 級表現
Sakana AI 的新編排模型在 DeepSWE、Chartography 等高難度基準上登頂，代理池中完全沒有 Fable 5 或 GPT-6 Astra——只靠可替換的開源與專用模型群，而 Fugu Max 的價格更比前沿模型低 40-60%。
-
Tools ENA Neck, Cooler Motors and a $14,000 Tag: Unitree's G1+ Makes Humanoids Practical
Unitree's G1+ adds a 2-DOF neck, 110% more shoulder torque, 72% cooler motors and far-field voice to its $14,000 humanoid — the platform refines itself toward real work.
-
Tools EN會轉頭、更耐熱、只要 1.4 萬美元：宇樹 G1+ 讓人形機器人走向實用
宇樹科技發表 G1+ 人形機器人：新增 2 自由度頸部、肩部扭矩提升 110%、發熱降低 72%，加上遠場語音互動，標準版售價 9.5 萬人民幣（約 1.4 萬美元）。
-
Models ENOpen Weights Strike Back: Nari Labs' Qwen3 Voice Endpoints Top Coval's Live Leaderboards
An eleven-person startup built an inference engine that runs Alibaba's open Qwen3-TTS/ASR faster and cheaper than ElevenLabs and Deepgram — and open-sourced the recipe.
-
Models 中開源權重的逆襲：Nari Labs 的 Qwen3 語音端點登上 Coval 即時排行榜頂端
一家約十人的新創打造了專用推論引擎，讓阿里巴巴開源的 Qwen3-TTS/ASR 跑得比 ElevenLabs 和 Deepgram 更快、更便宜——而且把整套方法開源。
-
Models ENThe Quiet Workhorse: Kimi K2.8 Preview Replaces kimi-for-coding Overnight
Moonshot upgraded its default coding model in place — K2.8 Preview now serves every kimi-for-coding request with near-K3 performance, 1M context, and cheaper thinking.
-
Models 中沉默的工作馬：Kimi K2.8 Preview 一夜之間接管 kimi-for-coding
Moonshot 低調完成原地換引擎——K2.8 Preview 全面接管 kimi-for-coding 的所有請求，帶來接近 K3 的效能、百萬 token 上下文，以及更便宜的推理。
-
Industry ENPortrait Mode's Creators Join OpenAI: The Glass Imaging Acquisition Decoded
OpenAI quietly acquired Glass Imaging, the startup founded by the ex-Apple engineers behind iPhone Portrait Mode, for over $300 million — a signal that its screen-free hardware push needs world-class cameras.
-
Industry 中人像模式推手加入 OpenAI：解讀 Glass Imaging 收購案
OpenAI 低調收購了由 iPhone 人像模式背後前 Apple 工程師創辦的 Glass Imaging，金額超過 3 億美元——這釋出的訊號是：無螢幕硬體裝置需要世界級的相機技術。
-
Tools ENFrom 150,000 Physical Qubits to 1,000 Logical: NVIDIA's CUDA-Q Logical Aims to Industrialize Fault-Tolerant Quantum Design
NVIDIA's new CUDA-Q Logical orchestration layer lets researchers codesign fault-tolerant quantum systems in software — Fermilab cut a five-month design cycle to three weeks, and Iceberg Quantum found a path to 1,000 logical qubits with 10x fewer physical qubits.
-
Tools 中從 15 萬顆實體量子位元到 1,000 顆邏輯量子位元：NVIDIA CUDA-Q Logical 要把容錯量子電腦的設計變成軟體工程
NVIDIA 開源全新 CUDA-Q Logical 編排層，讓研究人員在軟體中共同設計容錯量子系統——費米實驗室把五個月的設計週期壓縮到三週，Iceberg Quantum 更找到用十分之一實體量子位元做出 1,000 顆邏輯量子位元的路徑。
-
Industry ENThe Fabric Learns to Compute: Cornelis Unveils Active Compute Fabric, a $205M Raise and a Qualcomm Alliance
Cornelis Networks enters scale-up networking with an open, programmable fabric backed by $205M in new funding and a Qualcomm alliance — aiming at the ~50% of GPU hours giant AI clusters lose to waiting for data.
-
Industry 中讓網路織網學會運算：Cornelis 發表 Active Compute Fabric、2.05 億美元融資與 Qualcomm 結盟
Cornelis Networks 以開放、可編程的網路架構進軍 scale-up 連網市場，搭配 2.05 億美元融資與 Qualcomm 合作——瞄準大型 AI 叢集中約半數 GPU 時數浪費在等待資料的痛點。
-
Industry ENTomorrow the Web's Default Setting Flips: Cloudflare's Content Independence Day Arrives
On September 15, Cloudflare blocks Training and Agent crawlers by default on ad-bearing pages for all new domains — the largest single shift in crawler economics ever attempted, and a wake-up call for mixed-use bots like Googlebot.
-
Industry 中明天，網路的預設值翻轉：Cloudflare「內容獨立日」正式到來
9 月 15 日起，Cloudflare 對所有新接入網域的廣告頁面預設封鎖 Training 與 Agent 爬蟲——這是爬蟲經濟史上最大規模的一次預設值翻轉，也是 Googlebot 等混合用途爬蟲的最大警訊。
-
Industry ENEurope's Biggest Round Ever: Mistral Raises €3B Led by Samsung to Make Open-Weight AI the Frontier
Samsung-led €3B Series D values the French lab above €21B — the largest equity round in European tech history — with co-leads from EQT's EU-backed Scaleup Europe Fund and PSG Equity.
-
Industry 中歐洲史上最大募資：Mistral 獲三星領投 30 億歐元 D 輪，要讓開放權重 AI 站上前線
三星領投的 30 億歐元 D 輪將這家法國實驗室估值推上 210 億歐元以上，創歐洲科技公司股權募資紀錄，EQT 管理的歐盟 Scaleup Europe Fund 與 PSG Equity 共同領投。
-
Industry ENNegative-Yield Convexity: Why Hong Kong Just Handed Z.AI $5 Billion for Free
Z.AI priced $3B of zero-coupon convertible bonds at a negative yield and placed $2B of stock in one move — the most aggressive capital raise in the LLM industry's short history.
-
Industry 中負殖利率的豪賭：香港為何免費奉上 50 億美元給 Z.AI
Z.AI 一口氣完成 20 億美元配股加上 30 億美元零息可轉債——定價隱含負殖利率——這是 LLM 產業短短歷史中最激進的一次募資。
-
Models ENNamed Before It Ships: Musk Reveals Grok 4.8, a 2.5T Model on a Rewritten C++ Stack, While 4.7 Is Still Missing
In a single reply post, Elon Musk named Grok 4.8 — 2.5 trillion parameters, trained on SpaceXAI's rewritten C++ stack — while Grok 4.7 remains unreleased past its fourth missed date. We trace the four-month paper trail behind the claim.
-
Models EN先命名、後出貨：Musk 揭曉 Grok 4.8——2.5 兆參數、全新 C++ 軟體堆疊，而 4.7 依然缺席
馬斯克在一則回覆貼文中命名了 Grok 4.8：2.5 兆參數、以 SpaceXAI 重寫的 C++ 堆疊訓練。此時 Grok 4.7 已錯過第四個期限、仍未發布。我們追溯這項宣告背後長達四個月的線索。
-
Models ENVision as a Feedback Loop, Not Just an Input: Ant Group Open-Sources the 124B Ling-3.0-flash-VL Under MIT
Ant Group's inclusionAI has open-sourced Ling-3.0-flash-VL, a 124B-parameter multimodal MoE that activates only 5.5B per token, reads images and video over a 256K context, ships in BF16 and FP8, and carries a permissive MIT license.
-
Models 中視覺是回饋迴路，而不只是輸入：螞蟻集團以 MIT 授權開源 124B 的 Ling-3.0-flash-VL
螞蟻集團旗下 inclusionAI 開源 Ling-3.0-flash-VL：124B 參數的多模態 MoE 模型，每 token 僅啟動 5.5B 參數，支援影像與影片輸入、256K 上下文，提供 BF16 與 FP8 權重，並採用寬鬆的 MIT 授權。
-
Industry ENFrom $26B to $48B in Four Months: Cognition's $2 Billion Series E Says AI Coding Is Now a Category War
Devin-maker Cognition raised over $2 billion at a $48 billion valuation, led by new investors a16z and Accel, as run-rate revenue nearly doubled to $900 million in four months and investors bet the coding-agent market has room for multiple winners.
-
Industry 中四個月估值從 260 億翻至 480 億美元：Cognition 的 20 億美元 Series E 宣告 AI 編碼進入多強割據時代
Devin 開發商 Cognition 宣布完成超過 20 億美元 Series E 融資，估值達 480 億美元，由新投資人 a16z 與 Accel 領投；年化營收四個月內近乎翻倍至 9 億美元，投資人押注編碼代理市場容得下多個贏家。
-
Industry ENFive Initiatives, One Open-Source Pitch: Xi Takes China's AI Vision to BRICS in Delhi
At the 18th BRICS Summit in New Delhi, Xi Jinping proposed a China-led BRICS AI open-source community, pitched membership in Beijing's World AI Cooperation Organization, and made open weights the diplomatic currency of the Global South.
-
Industry 中五項倡議、一場開源提案：習近平把中國的 AI 願景帶進德里 BRICS 峰會
在第 18 屆 BRICS 新德里峰會上，習近平提出由中國主導的 BRICS AI 開源社群，邀請各國加入北京的世界人工智慧合作組織，讓開放權重成為全球南方的外交貨幣。
-
Models ENThe Sunset That Wasn't: DeepSeek V4.1-Flash Ships With 1M Context and $0.003 Cache Reads — and V4 Pro Gets a Reprieve
DeepSeek's new 552B-parameter open-weight model undercuts rivals on inference pricing just as the company walks back its plan to retire the V4 Pro API on September 14.
-
Models 中一場沒有發生的日落：DeepSeek V4.1-Flash 以百萬級上下文與 0.003 美元快取讀取登場，V4 Pro 獲得緩刑
DeepSeek 這款 552B 參數的開放權重新模型以極低的推理價格挑戰對手，同時公司收回於 9 月 14 日關閉 V4 Pro API 的計畫。
-
Industry ENSqueezing Tokens: How Chinese AI Labs Are Closing the Gap by Being More Efficient
A new Fortune analysis argues China's labs are matching US frontier AI not by outspending it but by out-engineering it — cheaper attention algorithms, open weights, and enterprise adoption data that show the gap has narrowed to 2.7%.
-
Industry 中把 Token 榨到極致：中國 AI 實驗室靠「效率」追平美國的真相
Fortune 九月十三日分析指出，中國實驗室並非只靠蒸餾美國模型——在算力封鎖下被迫練出的效率工程，才是它們把差距縮到 2.7%、並開始反過來拿下美國企業客戶的真正武器。
-
Research EN447 Papers in One Day: A CMU Professor Says CS Academia Should 'Burn to the Ground'
As arXiv's machine-learning category logs a record 447 submissions in a single day, CMU's Zachary Lipton declares that 'CS academia broke the system' — and perhaps it must burn to rebuild. Inside the numbers behind a researcher's breaking point.
-
Research 中單日 447 篇論文：CMU 教授說資科學界應該「燒掉重來」
arXiv 機器學習分類單日湧入 447 篇新論文創下紀錄之際，CMU 的 Zachary Lipton 直言「資科學界弄壞了這套系統」——或許唯有燒掉才能重建。一位研究者崩潰宣言背後的數字真相。
-
Policy EN'Stop Pretending This Is Altruism': Sacks Calls Amodei's Pacing Plan 'Regulatory Capture' as the Backlash Arrives
The day after Musk, Altman, and Hugging Face endorsed Dario Amodei's 'We Must Pace the Frontier,' the backlash landed: White House AI czar David Sacks branded the plan 'regulatory capture,' Meta's Alexandr Wang said alignment could become 'the gating factor for scaling,' and Musk insisted 'nothing can shut down open source.'
-
Policy 中「別再假裝這是利他主義」：Sacks 痛批 Amodei 放緩計畫是「監管俘虜」，反彈浪潮正式登場
在 Musk、Altman 與 Hugging Face 表態支持 Amodei 的《我們必須為前沿放緩腳步》之後，反彈接踵而至：白宮 AI 沙皇 David Sacks 直斥該計畫是「監管俘虜」，Meta 的 Alexandr Wang 稱對齊可能成為「規模化的瓶頸因子」，Musk 則堅持「沒有什麼能關掉開源」。
-
Tools ENAn Inbox for Your Agents: AWS Open-Sources Pizza Bot, Born From 2,000 Internal Users
AWS has open-sourced Pizza Bot, a self-hosted, Apache 2.0 inbox for background AI agents built on DeepAgents and LangGraph — email metaphors, durable approvals, cron and webhooks, and a model-provider menu from Bedrock to Ollama.
-
Tools 中給代理一個收件匣：AWS 開源 Pizza Bot，從 2,000 名內部用戶長出來的背景代理工作台
AWS 開源了 Pizza Bot：一套自架、Apache 2.0 授權的背景 AI 代理收件匣，以 DeepAgents 與 LangGraph 构建，用電子郵件的隱喻、可持久暫停的審批、cron 與 webhook 觸發，支援從 Bedrock 到 Ollama 的多家模型供應商。
-
Meta ENYour Code Editor Is Phoning Home: huggingface_hub Silently Fingerprints 26 AI Coding Agents
A network-traffic audit found the huggingface_hub SDK scanning environment variables for 26 coding agents — from Claude Code and Cursor to Warp and Zed — and tagging every Hub request with an agent/<name> user-agent. Hugging Face publishes the aggregate numbers monthly; most developers never knew the telemetry existed.
-
Meta 中你的編輯器正在回報身分：huggingface_hub 低調指紋辨識 26 款 AI 編碼代理
一份網路流量稽核發現，huggingface_hub SDK 會掃描環境變數以辨識 26 款編碼代理——從 Claude Code、Cursor 到 Warp 與 Zed——並在每次 Hub 請求加上 agent/<名稱> 的 user-agent 標記。Hugging Face 每月公開彙整數據，但多數開發者根本不知道這套遙測存在。
-
Research ENDOOMFLY: A Complete Fruit Fly Brain, Simulated and Trained to Play Doom
Days after researchers published the complete MaleCNS v1.0 connectome, a Coinbase engineer wired 166,700 simulated fly neurons into Doom — with damage feeding a two-cell dopamine loop — and the internet followed with Mario 64 and Beat Saber.
-
Research 中DOOMFLY：完整的果蠅大腦被模擬出來，然後被訓練去玩《毀滅戰士》
MaleCNS v1.0 連結體發布僅十天後，一位 Coinbase 工程師就把 16.67 萬顆模擬果蠅神經元接上了《毀滅戰士》——受傷訊號直接餵進兩顆真實的多巴胺細胞——隨後網友又讓它玩起了《瑪利歐 64》和《Beat Saber》。
-
Models ENThree Open Weights Against the API: Inside Abacus.AI's Smaug Line for Enterprise Agents
Abacus.AI's new Smaug Agentic, Flash, and Mini fine-tunes of Kimi K3, DeepSeek V4 Flash, and Qwen3.8 27B beat Claude Sonnet 5 on several agentic benchmarks — with weights you can download.
-
Models 中三組開放權重對決封閉 API：深入 Abacus.AI 為企業 Agent 而生的 Smaug 模型家族
Abacus.AI 新推出的 Smaug Agentic、Flash 與 Mini，分別微調自 Kimi K3、DeepSeek V4 Flash 與 Qwen3.8 27B，在多項 Agent 基準測試上超越 Claude Sonnet 5——而且權重可以直接下載。
-
Research ENReal-SWE: Coding Agents Graded on Code No Model Has Ever Seen — Fable 5.1 Leads at 38.8%
Specific Labs' new Real-SWE benchmark tests coding agents on licensed private enterprise codebases no model could have memorized. Fable 5.1 resolves 38.8% of tasks, GPT-6 Astra 33.8%, and open-weight GLM-5.3 surprises at 28.8% — but six of ten sample tasks sit below a 15% resolution rate.
-
Research 中Real-SWE：在沒有任何模型見過的程式碼上評測 Coding Agent——Fable 5.1 以 38.8% 奪冠
Specific Labs 推出的 Real-SWE 基準測試，在授權取得的私有企業程式碼庫上評測 coding agent，杜絕訓練資料記憶作弊。Fable 5.1 解決率 38.8%、GPT-6 Astra 33.8%，開放權重的 GLM-5.3 意外以 28.8% 緊咬——但十項樣本任務中有六項解決率低於 15%。
-
Industry ENThe $1.5 Billion Hire: Google Completes Its Mechanize Talent Deal
LinkedIn profiles confirm Google has closed its reported $1.5B+ talent-and-license deal with Mechanize, the 35-person RL-environment startup founded by Epoch AI alumni to automate all work — a acqui-hire priced like an acquisition.
-
Industry 中15 億美元的「錄用」：Google 完成與 Mechanize 的人才交易
LinkedIn 資料證實 Google 已完成與 Mechanize 金額超過 15 億美元的人才加授權交易。這家由 Epoch AI 研究員創立、以「全面自動化經濟」為使命的 35 人新創，以收購等級的身價完成了一場結構上只是「大規模錄用」的出走。
-
Industry ENThey Found Their Own Code in Google's ARTEMIS — and a Force-Push That Erased Their Names
Paris startup Minitap says Google's Pixel team copied its Apache 2.0 mobile-use repo into ARTEMIS, then force-pushed the authors' names out of the package file. An open-source attribution fight with implications for every maintainer.
-
Industry 中他們在 Google ARTEMIS 裡發現自己的程式碼——還有一個抹掉他們名字的 Force-Push
巴黎新創 Minitap 指控 Google Pixel 團隊將其 Apache 2.0 授權的 mobile-use 程式碼整段搬進 ARTEMIS，再透過 force-push 把原作者姓名從套件檔案中移除。這場開源歸屬權之爭，關乎每一位維護者。
-
Meta ENWithin the Rules: SGLang's SafeUnpickler Bypass Is the Fourth Critical AI-Infra CVE in 18 Days
CVE-2026-86793 lets an unauthenticated attacker run arbitrary code on SGLang inference servers by chaining two permitted builtins functions — no patch exists yet. It is the fourth critical CVE to hit the AI stack since August 25.
-
Meta 中在規則之內繞過規則：SGLang SafeUnpickler 旁路成為 18 天內第四個關鍵 AI 基礎設施 CVE
CVE-2026-86793 讓未經身分驗證的攻擊者只需串接兩個「被允許」的 Python 內建函式，就能在 SGLang 推論伺服器上執行任意程式碼，且官方修補至今尚未推出。這是 8 月 25 日以來第四個重創 AI 技術棧的關鍵 CVE。
-
Models ENOne GPU, 262K Context, Apache 2.0: Agnes-3.0-Flash Makes Hybrid-Attention Multimodal Reasoning a Single-Card Affair
Agnes-AI open-sources a 33B hybrid-attention multimodal model with a 262,144-token context, image and video understanding, and 85.05 GPQA Diamond — all on one H100 at bf16.
-
Models 中單卡跑 262K 上下文的 Apache 2.0 多模態模型：Agnes-3.0-Flash 讓混合注意力推理走進單 GPU 時代
Agnes-AI 開源 33B 混合注意力多模態模型，支援 262,144 token 上下文、圖片與影片理解，GPQA Diamond 拿下 85.05——bf16 下單張 H100 即可部署。
-
Models ENTen Days of Hype, Then a Hold: Musk Delays Grok 4.7 Hours Before Launch to Fix an RL Bug
Grok 4.7 — the 2.1T-parameter model trained on SpaceX data — was due September 12. Instead, Musk delayed it 'a few days' after RL training penalized long answers and made the model give up on hard tasks.
-
Models EN十天造勢之後按下暫停：Musk 在 Grok 4.7 發表前幾小時宣布延期，只為修一個 RL 獎勵漏洞
原定 9 月 12 日登場、以 SpaceX 資料訓練的 2.1 兆參數模型 Grok 4.7 臨時延期。原因：RL 訓練對回應長度的懲罰過重，導致模型在困難任務上直接放棄。
-
Research ENTwo Million Moon Tiles and 1,100 GPU-Hours: Inside NASA and IBM's Open-Source Lunar Foundation Model
NASA and IBM have open-sourced a multimodal lunar foundation model trained on SomBench, a 39TB corpus spanning 17 years of LRO observations — cutting polar ice-mapping error by up to 22% while beating ImageNet baselines with half the training data.
-
Research 中兩百萬張月球影像、1,100 GPU 小時：NASA 與 IBM 開源月球基礎模型全解析
NASA 與 IBM 開源了以 SomBench（39TB、涵蓋 17 年 LRO 觀測）訓練的多模態月球基礎模型——極區冰層預測誤差最多降低 22%，且僅用一半訓練資料就擊敗 ImageNet 基線。
-
Models EN8B Active Parameters and a 890-Byte Cache: DeepSeek's V4.1-Flash Outscores Its Own Flagship — and Phases It Out
DeepSeek's new 552B-parameter MoE activates just 8B parameters per input token, compresses its KV cache to 890 bytes per token, beats V4-Pro on Terminal-Bench 2.1 and DeepSWE — and inherits the flagship's API traffic on September 14.
-
Models EN每 token 僅啟動 8B 參數、KV 快取壓到 890 位元組：DeepSeek V4.1-Flash 全面超越自家旗艦——然後把它淘汰
DeepSeek 新發布的 552B 參數 MoE 模型，每個輸入 token 僅啟動 8B 參數，KV 快取壓縮至每 token 890 位元組，在 Terminal-Bench 2.1 與 DeepSWE 上擊敗 V4-Pro——並將於 9 月 14 日承接旗艦模型的 API 流量。
-
Industry ENFrom Demos to $100M in Ten Months: Skild AI Declares the Era of Robot Deployment Open
Skild AI has crossed $100M in annual revenue run rate just ten months after its first commercial deployment — 60+ customers, hundreds of robots, and a generalist brain that learns new tasks from a single video.
-
Industry 中從 Demo 到 1 億美元只花十個月：Skild AI 宣告機器人部署時代來臨
Skild AI 在首次商業部署後十個月達成 1 億美元年營收跑速率——超過 60 家客戶、數百台機器人，靠的是一支影片就能學會新任務的通用機器人大腦。
-
Industry ENOttawa Writes a Cheque: Cohere Nears $2-3 Billion Round at $20 Billion, in What Would Be Canada's Largest Private Startup Deal
Cohere is in advanced talks to raise US$2-3 billion at a US$20 billion valuation with the Canadian government investing and Berlin in talks — set to become the largest private financing by a Canadian startup on record.
-
Industry 中渥太華出手投資：Cohere 洽談以 200 億美元估值募資 20-30 億，可望寫下加拿大新創私募紀錄
總部位於多倫多的 Cohere 正深入洽談以 200 億美元估值募資 20 至 30 億美元，加拿大政府確定參與、德國政府也在談——一旦成交將是加拿大新創公司史上最大私募輪。
-
Industry ENWafer to Token: Nvidia Puts Palantir's Sovereign AI Stack in Charge of Its 1.3M-Part Supply Chain
Announced at AIPCon 11, the Nvidia-Palantir stack fuses Nemotron open models, cuOpt and the Palantir Ontology into a governed learning loop — deployed first inside Nvidia's own 'wafer to first token' supply chain, with Dell, Cisco, Rackspace and Nebius as infrastructure partners.
-
Industry 中從晶圓到 Token：Nvidia 用 Palantir 主權 AI 技術棧管理 130 萬零件供應鏈
在 AIPCon 11 發布的 Nvidia-Palantir 聯合技術棧，將 Nemotron 開源模型、cuOpt 與 Palantir Ontology 融入受治理的學習迴圈，率先部署於 Nvidia 自家「從晶圓到首個 token」的供應鏈，並由 Dell、Cisco、Rackspace 與 Nebius 擔任基礎設施夥伴。
-
Models ENTwo Models, One Frontier: Sakana's Fugu Max and Fugu Ultra v2 Bet That Orchestration Beats Monoliths
Sakana AI ships Fugu Max (frontier-grade scores at 40-60% lower output cost) and Fugu Ultra v2 (best on 5 of 8 benchmarks — without GPT-6-Astra or Fable 5 in the pool).
-
Models 中兩個模型、一條前緣：Sakana 的 Fugu Max 與 Fugu Ultra v2 押注「編排」擊敗巨型單體模型
Sakana AI 發布 Fugu Max（輸出價格比 Sonnet 5、GPT 5.6 Terra 低 40-60%）與 Fugu Ultra v2（8 項基準中 5 項奪冠——而且模型池裡根本沒有 GPT-6-Astra 或 Fable 5）。
-
Meta ENYou Thought You Were Talking to Kimi: Inside the Relay Scheme That Served Claude to Millions
Anthropic's threat report reveals Moonshot and DeepSeek silently forwarded real customer requests to Claude and showed the answers as their own — 151 million exchanges for Alibaba, cross-session replay attacks against encrypted reasoning, and developer secrets from a dozen countries caught in the middle.
-
Meta 中你以為在跟 Kimi 對話：起底把 Claude 偷偷送進數百萬用戶對話的中轉計畫
Anthropic 威脅報告揭露，Moonshot 與 DeepSeek 曾把真實客戶的請求悄悄轉發給 Claude，再把答案當成自家模型的回覆——Alibaba 量產 1.51 億次對話、跨 session 重放攻擊破解加密推理、十幾國開發者的商業機密全被捲入。
-
Industry ENFrom $300M to $1B in Two Months: Moonshot's Kimi K3 Just Rewrote the Open-Weight Business Model
Moonshot AI told investors its annualized revenue crossed $1B in August — up from $300M in June — driven by Kimi K3, the largest open-weight model ever released. It now targets $2B ARR by year-end ahead of a Hong Kong IPO.
-
Industry 中兩個月從 3 億到 10 億美元：Moonshot 的 Kimi K3 改寫了開放權重商業模式
Moonshot AI 向投資人透露，其年化營收在 8 月突破 10 億美元——較 6 月的 3 億美元大幅躍升，幕後推手是史上最大開放權重模型 Kimi K3。在香港 IPO 之前，公司目標年底達到 20 億美元 ARR。
-
Research EN30 Out of 42, Fully Open-Sourced: NVIDIA Publishes the Complete Recipe Behind Nemotron's IMO Gold
NVIDIA releases the checkpoints, training data, inference code, submitted solutions and a 200-problem benchmark behind its natural-language pipeline that scored IMO-gold 30/42 with no formal prover.
-
Research 中42 題拿 30 分、全程開源：NVIDIA 公開 Nemotron IMO 金牌背後的完整配方
NVIDIA 釋出自然語言數學證明管線的全部成果——兩個後訓練檢查點、訓練資料、推理程式碼、競賽實際提交的解答，以及 200 題全新奧林匹亞基準，在不依賴形式證明器的條件下達到 IMO 金牌門檻 30/42。
-
Models ENThe Model That Pretends to Be You: humans& Ships Persimmon, a 550B User Simulator
humans& released Persimmon v0.1, a 550B-parameter model post-trained from NVIDIA's Nemotron 3 Ultra whose job is to convincingly simulate human users — fooling LLM judges 18.6–21.1% on Multi-User Turing tests versus under 3% for frontier assistants.
-
Models 中那個假扮成你的模型：humans& 發布 550B 參數「使用者模擬器」Persimmon
humans& 於 9 月 10 日發布 Persimmon v0.1——以 NVIDIA Nemotron 3 Ultra 為基礎後訓練的 550B 參數模型，唯一任務是逼真模擬人類使用者，在 Multi-User 測靈測試中騙過 LLM 評審的比率達 18.6–21.1%，遠高於前沿助理模型的不到 3%。
-
Models EN83.6 on WMT26 With 25B Active Parameters: Cohere's North Small Translate Outscores DeepL, Google — and Gives the Weights Away
Cohere's first dedicated translation model — a 218B-parameter MoE with only 25B active — posts 83.60 on WMT26 All Languages, beating DeepL NextGen (81.37), Qwen 3.5 397B (81.56) and Google Translate (68.20), runs on two H100s, and ships as open weights.
-
Models 中25B 活躍參數拿下 WMT26 83.6 分：Cohere 北極星小翻譯模型擊敗 DeepL 與 Google，還直接開源權重
Cohere 首款專用翻譯模型 North Small Translate 以 218B 總參數、僅 25B 活躍參數的 MoE 架構，在 WMT26 全語言基準拿下 83.60 分，超越 DeepL NextGen（81.37）、Qwen 3.5 397B（81.56）與 Google 翻譯（68.20），兩張 H100 就能跑，而且開放權重下載。
-
Models ENBeats Suno v5 on SongBench and Runs on a 24GB GPU: m-a-p's YuE2 Makes Open Music Generation Frontier-Grade
The open-source m-a-p collective just shipped YuE2-3B, an open-weight music model that tops WildSongBench with an editable ABC score, agentic editing, cover generation, and 71-second songs on an RTX 4090.
-
Models 中開源音樂模型站上前緣：m-a-p 的 YuE2 在 SongBench 擊敗 Suno v5，24GB 顯卡就能跑
開源社群 m-a-p 發布 YuE2-3B 開放權重音樂模型：WildSongBench 綜合分超越所有閉源系統，支援可編輯 ABC 樂譜、Agent 式改歌與翻唱，RTX 4090 上 71 秒生成一首歌。
-
Models EN8B Open Weights Beat the Closed Frontier: SenseTime's SenseNova-U1.5 Outscores Nano-Banana-Pro on Vision Reasoning
SenseTime's 8B-parameter SenseNova-U1.5-8B-MoT natively unifies visual understanding and generation without encoders or VAEs — and posts 68.2% on VBVR-Pro-Bench, ahead of proprietary Nano-Banana-Pro at 56.4%, under Apache 2.0.
-
Models 中80 億參數開源擊敗閉源前沿：商湯 SenseNova-U1.5 視覺推理評測超越 Nano-Banana-Pro
商湯以 Apache 2.0 釋出的 8B 原生統一多模態模型 SenseNova-U1.5-8B-MoT，在 VBVR-Pro-Bench 視覺推理評測拿下 68.2%，超越專有模型 Nano-Banana-Pro 的 56.4%，並原生支援 4K 生成。
-
Models EN64% Cheaper, One Point Shy of the Frontier: Cognition's SWE-2 Trains Every Effort Level in a Single RL Run
Cognition's SWE-2, post-trained from Moonshot's open-weight Kimi K3, scores 50.0% on FrontierCode 1.1 Main — within a point of Anthropic's Fable 5.1 — at 64% lower cost, using a Pareto-informed cost penalty to train medium, high and max effort levels in one RL run.
-
Models 中便宜 64%、距離前沿只差一分：Cognition SWE-2 用單次 RL 訓練搞定所有推理力度
Cognition 以 Moonshot 開源權重的 Kimi K3 為基底後訓練出的 SWE-2，在 FrontierCode 1.1 Main 拿下 50.0%，僅落後 Anthropic Fable 5.1 一分，成本卻低了 64%；其關鍵在於以 Pareto 導向的成本懲罰，在單次 RL 執行中同時訓練 medium、high 與 max 三種推理力度。
-
Industry EN€3 Billion and One Flag: Samsung Leads Mistral's Series D as Europe Buys Its Own AI Stack
Mistral AI has closed a €3 billion Series D at a €21+ billion valuation — the largest equity round in European tech history — to build sovereign, open-weight AI from models to data centers.
-
Industry 中30 億歐元一面旗：三星領投 Mistral D 輪，歐洲買下自己的 AI 全套技術棧
Mistral AI 完成 30 億歐元 D 輪融資、估值超過 210 億歐元——歐洲科技史上最大股權融資——要從模型到資料中心打造主權開源權重 AI。
-
Research ENFrom Petabytes to the Pale Moon: NASA and IBM Open-Source a Foundation Model for Lunar Science
Trained on 17 years of Lunar Reconnaissance Orbiter imagery, the NASA-IBM Lunar Foundation Model maps craters, spots young volcanism, and estimates polar ice stability — and it's free on Hugging Face.
-
Research 中從 PB 級資料到月球科學：NASA 與 IBM 開源專為月球打造的基礎模型
以月球勘測軌道器 17 年的影像資料訓練而成，NASA-IBM 月球基礎模型能繪製撞擊坑、辨識年輕火山特徵、估算極區冰層穩定性，現已在 Hugging Face 上免費開放。
-
Meta ENOne Command to Freedom: CVE-2026-82533 Let DeepSeek's 215k-Star Coding Agent Turn Off Its Own Sandbox
OX Research found that DeepSeek Harness (dsh) trusted the client-supplied Host header to gate its unauthenticated local API, so a sandboxed agent could elevate itself to danger-full-access and disable approval prompts with a single curl — shipped defaults, no credentials, no network exposure.
-
Meta 中一條指令逃出沙箱：CVE-2026-82533 讓 DeepSeek 215k 星標的編碼代理人親手關掉自己的牢籠
OX Research 發現 DeepSeek Harness（dsh）僅憑客戶端自行填寫的 Host 標頭來信任其未驗證的本機 API，沙箱內的代理人只要一條 curl 就能把自己升級成 danger-full-access 並關閉核准提示——出廠預設、無需憑證、無需對外曝露。
-
Models EN124B Parameters, Open Weights: Ant Group Open-Sources Ling-3.0-flash-Fin and the FinFIRST Benchmark
Ant Group releases Ling-3.0-flash-Fin, a finance-tuned MoE model with 124B total / 5.1B active parameters plus the expert-built FinFIRST benchmark — betting that Wall Street's next research analyst runs on open weights with auditable evidence chains.
-
Models 中1,240 億參數、開放權重：螞蟻集團開源 Ling-3.0-flash-Fin 與 FinFIRST 金融代理人基準
螞蟻集團開源金融強化模型 Ling-3.0-flash-Fin（124B 總參數／每 token 僅啟動 5.1B 的 MoE 架構），連同與中金公司投行團隊共同打造的專家級基準 FinFIRST——押注華爾街下一代研究分析師，將運行在可稽核證據鏈的開放權重模型上。
-
Models ENThe Flash That Ate the Flagship: DeepSeek V4.1 Flash Arrives — and Retires V4 Pro
DeepSeek's new 552B MoE with native multimodality beats its own flagship on agent benchmarks — so the company is routing all V4 Pro traffic to it, at Flash prices, from September 14.
-
Models 中吞噬旗艦的 Flash：DeepSeek V4.1 Flash 正式登場——並讓 V4 Pro 退役
DeepSeek 這款 552B 參數、原生多模態的新 MoE 模型在 agent 基準測試上擊敗自家旗艦——公司因此宣布 9 月 14 日起所有 V4 Pro 流量改由它承接，並以 Flash 價格計費。
-
Research EN36.5% and Nowhere Left to Hide: Φ-Bench Asks If Frontier LLMs Can Engineer the Infrastructure That Powers Them
A new 85-task benchmark from USTC, StepFun and Yale grounds LLM agents in real GPU kernel, training and serving codebases — the best model, Claude Opus 5, scores just 36.53%, and on hardware-edge tasks the field collapses to 5.4%.
-
Research 中36.5% 無所遁形：Φ-Bench 檢驗前沿 LLM 能否親手打造驅動自己的基礎設施
來自中科大、StepFun 與耶魯的 13 人團隊發布 85 項任務的新基準，讓 LLM 代理深入真實的 GPU 核心、訓練與推理程式碼庫——最強的 Claude Opus 5 僅拿 36.53%，硬體邊緣任務更是全場崩盤至 5.4%。
-
Tools ENFive Models, One Timeline: Adobe Turns Premiere Into the First Model-Agnostic AI Video Workbench
At IBC 2026, Adobe ships in-timeline generative video with five competing models — Firefly, Google Veo, Runway, Kling, and Luma — selectable from one toolbar, killing the export-import loop and quietly conceding Firefly Video has lost the model race.
-
Tools 中五個模型、一條時間軸：Adobe 把 Premiere 變成第一個模型中立的 AI 影片工作台
在 IBC 2026，Adobe 把生成式影片直接搬進 Premiere 時間軸，Firefly、Google Veo、Runway、Kling、Luma 五個彼此競爭的模型可在同一工具列切換，終結匯出再匯入的惡夢——也等於承認 Firefly Video 在模型競賽中已經落後。
-
Research ENData Beat Architecture: The Six-Year Ablation That Rewrites Pretraining History
A granular 6-year ablation by Dwarkesh Patel and Jerry Han finds data improvements delivered 3.24x more compute-efficiency gains than model architecture from 2019 to 2025 — 12.0x for data versus 3.7x for models — with the two axes almost perfectly independent.
-
Research 中資料擊敗了架構：一項六年消融實驗改寫預訓練史
Dwarkesh Patel 與 Jerry Han 發布縝密的六年消融實驗：2019 至 2025 年間，資料改進帶來的算力效率增益是模型架構的 3.24 倍——資料 12.0 倍、模型僅 3.7 倍——而且兩者幾乎完全獨立、互不依賴。
-
Industry ENThe Board Coup at WordPress: Automattic Puts Matt Mullenweg on Leave, CFO Takes Over
Automattic's board voted to place co-founder Matt Mullenweg on paid leave against his will, naming CFO Mark Davies interim CEO — the biggest shake-up in WordPress history.
-
Industry 中WordPress 王朝的董事會政變：Automattic 將 Matt Mullenweg 停職，CFO Mark Davies 暫掌執行長
Automattic 董事會表決將共同創辦人 Matt Mullenweg 強制帶薪停職，並指派 CFO Mark Davies 出任臨時執行長 — WordPress 歷史上最重大的一次權力重組。
-
Industry ENAgentic AI Moves to Orbit: Macron Unveils $1 Billion Altair-Next Gen, the World's Largest AI Satellite Infrastructure
France and the UAE are putting reasoning AI in space: a $1B, 50-satellite constellation with Mistral models onboard, alerts in seconds instead of hours, and an AI app store running on shared satellites.
-
Industry 中Agentic AI 進駐軌道：馬克宏揭幕 10 億美元 Altair-Next Gen，全球最大 AI 衛星基礎設施
法國與阿聯酋聯手把會推理的 AI 送上太空：10 億美元、50 顆衛星的星座，搭載 Mistral 模型在軌推理，警報從數小時縮短到數秒，並以 AI 應用商店模式讓多個模型共用同一批衛星。
-
Research ENAsk Devin: One Researcher and an Agent Swarm Just Factored RSA-260 and Made Breaking RSA 10x Cheaper
Cognition's Eric Lu drove up to 18 concurrent Devin sessions to build the world's fastest GPU lattice siever, factoring the 862-bit RSA-260 challenge number in 15 days on spare cluster compute for roughly $400,000 — and putting RSA-1024 within reach of any well-funded lab for about $30 million.
-
Research 中問 Devin 就對了：一位研究員帶 Agent 軍團分解 RSA-260，讓破解 RSA 的成本降為十分之一
Cognition 研究員 Eric Lu 驅動最多 18 個並行的 Devin session，打造出全球最快的 GPU 格篩法實作，在閒置叢集算力上花約 40 萬美元、15 天分解了 862 位元的 RSA-260 挑戰數——並讓任何資金充裕的實驗室都能以約 3,000 萬美元分解 RSA-1024。
-
Industry ENCheaper Than Open Source: OpenAI Pitches AI for Chip Design and Claims It Undercuts China on Price
At Goldman Sachs' Communacopia conference, CFO Sarah Friar said OpenAI's 80% Luna price cut drove 10x usage — and that its API can now beat Chinese open-source models on total cost in specialized verticals like chip design and life sciences.
-
Industry 中比開源還便宜？OpenAI 把 AI 推進晶片設計，宣稱在價格上擊敗中國對手
在高盛 Communacopia 科技會議上，OpenAI 財務長 Sarah Friar 表示 Luna 降價 80% 帶來十倍用量成長，並宣稱在晶片設計、生命科學等專業領域，其 API 的總持有成本已低於中國開源模型。
-
Meta EN100,000 Homegrown GPUs: JD Cloud and Moore Threads to Build China's Largest Domestic-Chip AI Cluster
At its Global Technology Explorers Conference, JD Cloud committed to a 100,000-GPU intelligent-computing cluster built entirely on Moore Threads' domestic full-function GPUs — the first time Chinese silicon reaches 100K scale inside a top AI cloud provider.
-
Meta EN十萬張國產 GPU：京東雲攜手摩爾線程，打造中國最大全國產晶片 AI 叢集
在 2026 京東全球科技探索者大會上，京東雲宣布擬建設十萬卡全功能 GPU 智算叢集，算力底座全面採用摩爾線程國產 GPU——這是國產晶片首次進入頭部 AI 雲廠商的十萬卡級核心叢集。
-
Tools EN18 Models, Zero Tokens: Desert Ant Labs Bets the Next AI Layer Runs on the Device, Not the Cloud
European lab Desert Ant Labs launched 18 small on-device models — 2-second transcription, 9MB studio audio, 12MB PII redaction — free under 100k monthly devices. HN scrutiny over what's under the hood only sharpened the story.
-
Tools 中18 個模型、零 token 成本：Desert Ant Labs 押注下一層 AI 在裝置上跑，不在雲端
歐洲新創 Desert Ant Labs 推出 18 個小型端側模型——2 秒轉錄、9MB 錄音室級降噪、12MB 個資遮蔽——10 萬月活裝置內免費。Hacker News 對其技術底細的猛烈檢驗，反而讓故事更清晰。
-
Industry EN$1.35 Billion for Milliwatts: Analog Devices Buys Alif Semiconductor to Put AI on a Battery
ADI will pay $1.35B in cash (up to $1.55B with earnouts) for Alif Semiconductor, whose AI-native microcontrollers and fusion processors run neural networks on milliwatts — the edge-computing counterweight to gigawatt data centers.
-
Industry 中13.5 億美元買毫瓦：Analog Devices 收購 Alif Semiconductor，要把 AI 裝進電池裝置
ADI 以 13.5 億美元現金（含業績獎金最高 15.5 億美元）收購 Alif Semiconductor，其 AI 原生微控制器與融合處理器能用毫瓦級功耗跑神經網路——這是 GW 級資料中心之外的另一端 AI 經濟。
-
Industry ENNvidia Buys the 'GitHub of AI': Inside the $12.93 Billion Hugging Face Deal
Nvidia's largest AI software acquisition puts the world's dominant open-model platform under the world's most valuable chipmaker — with 19 mentions of 'open' and a neutrality pledge Nvidia now has to keep.
-
Industry 中Nvidia 買下「AI 界的 GitHub」：129.3 億美元收購 Hugging Face 全解析
全球最強 AI 晶片商以 129.3 億美元收購最重要的開源模型平台，聲明稿裡出現 19 次「open」，但中立性承諾能否兌現，才是這筆交易真正的考題。
-
Models ENA Flash Tier Retiring the Pro: Inside DeepSeek's V4.1 Flash Beta and Its 48-Hour Deadline
DeepSeek quietly opened a two-day beta for V4.1 Flash — a new-architecture, natively multimodal model that claims to surpass V4 Pro at Flash pricing, with outputs peaking at 507 tokens/s.
-
Models 中用 Flash 退役 Pro：DeepSeek V4.1 Flash 限時 48 小時公開測試全解析
DeepSeek 低調開放 V4.1 Flash 兩天限時測試——全新架構、原生多模態，宣稱以 Flash 價格超越 V4 Pro，輸出速度實測最高達每秒 507 tokens。
-
Research EN49 Billion Compounds, Nine Finalists, One Approval: Inside Mprosevir, China's First AI-Assisted Class 1 Drug
China's NMPA has granted conditional approval to Mprosevir, an AI-assisted COVID-19 antiviral that went from candidate discovery to completed clinical trials in 3.5 years — the world's first approved small molecule born from DNA-encoded library technology.
-
Research EN49 億化合物、9 個入圍者、1 張藥證：解析中國首款 AI 輔助一級新藥 Mprosevir
中國國家藥監局有條件核准 Mprosevir——一款 AI 輔助開發的 COVID-19 抗病毒藥物，從候選藥物發現到臨床試驗完成僅花 3.5 年，也是全球首款源自 DNA 編碼文庫技術獲批的小分子新藥。
-
Policy ENFrom Pilot to Permanence: NSF's $35M NAIRR Operations Center Puts SDSC and TACC in Charge of America's AI On-Ramp
NSF has awarded $35 million over five years to SDSC and TACC to run the NAIRR Operations Center — the operational backbone meant to turn a successful pilot that served 800+ projects into permanent national AI infrastructure.
-
Policy 中從試點到常態：NSF 拍板 3,500 萬美元成立 NAIRR 營運中心，SDSC 與 TACC 接管全美 AI 研究資源入口
美國國家科學基金會（NSF）宣布以五年 3,500 萬美元的協議，委由聖地牙哥超級電腦中心（SDSC）與德州先進運算中心（TACC）成立 NAIRR 營運中心，把服務超過 800 個研究計畫的試點計畫，轉型為常態化的國家級 AI 基礎建設。
-
Models ENFrom Simulating the World to Acting in It: HiDream's O1-Embodied Tops RoboColiseum's Hardest Leaderboard
HiDream.ai launches HiDream-O1-Embodied, an embodied world model that ranks No. 1 on RoboColiseum's Robustness leaderboard (0.692) and pairs a 100x generative data flywheel with Noitom motion capture.
-
Models 中從模擬世界到親手操作：HiDream 推出 O1-Embodied，奪下 RoboColiseum 最難排行榜榜首
HiDream.ai 發布具身世界模型 HiDream-O1-Embodied，以 0.692 分登上 RoboColiseum 魯棒性排行榜第一名，並以 Noitom 動捕資料結合 100 倍生成式擴增打造資料飛輪。
-
Research ENDrinking Water Beside an Ocean: Terence Tao Warns That Good Math Problems Are Now a Non-Renewable Resource
Hours after OpenAI's contested Navier–Stokes announcement, Fields Medalist Terence Tao laid out the deeper worry in a four-part essay: without boundaries on AI solution-mining, the field's scarce resource is not answers but good questions — and the incentives now point toward secrecy.
-
Research 中被海洋包圍卻缺飲用水：Terence Tao 警告，好的數學難題正在成為「不可再生資源」
在 OpenAI 引發爭議的 Navier–Stokes 宣布數小時後，菲爾茲獎得主 Terence Tao 發表四部曲長文提出更深層的憂慮：若不對 AI 的「解題挖礦」設下界線，數學界最稀缺的資源將不是答案，而是好問題——而現在的誘因正把整個領域推向保密。
-
Industry ENSamsung Puts Mistral AI Inside the Fab: On-Premises LLMs for Every Chip Plant
Announced during the Korea–France summit in Paris, Samsung's strategic partnership with Mistral AI brings Mistral Large and customized on-premises models into semiconductor design and manufacturing — defect detection, equipment optimization, and yield stabilization across memory, logic, and foundry.
-
Industry 中三星把 Mistral AI 裝進晶圓廠：全面導入在地化 LLM 的智慧製造布局
在巴黎舉行的韓法高峰會上，三星電子宣布與 Mistral AI 建立策略夥伴關係，將 Mistral Large 與客製化在地部署模型導入半導體設計與製造——涵蓋記憶體、邏輯與晶圓代工的缺陷檢測、設備優化與良率穩定。
-
Research EN23.9% vs 82.2%: τ^τ-Bench Makes AI Build the Agents It Used to Only Answer For
A new 53-task benchmark hands coding agents the messy artifacts of a real client engagement and asks them to ship a working customer-service bot. The best one passes under a quarter of evaluations; expert-built references pass 82.2%.
-
Research 中23.9% 對 82.2%：τ^τ-Bench 讓 AI 從「當客服代理」升級去「蓋客服代理」
全新 53 任務基準測試把真實客戶委託案的雜亂素材整包丟給程式代理，要求它交付可上線的客服機器人。最強組合通過率不到四分之一，專家打造的參考實作則有 82.2%。
-
Industry EN40,000 Square Feet of Cleanroom for the Post-Silicon Era: HCLTech Opens a ₹185 Crore Advanced Semiconductor Lab in Bengaluru
HCLTech launched a 40,000 sq ft Advanced Semiconductor Lab in Bengaluru with ₹185 crore of investment, betting that India's next chip opportunity is not fabs but the unglamorous engineering that happens after silicon comes back from the foundry.
-
Industry 中為後矽晶世代而建的 4 萬平方英尺無塵室：HCLTech 於班加羅爾啟用 185 億盧比先進半導體實驗室
HCLTech 於班加羅爾啟用佔地 4 萬平方英尺、投資額 185 億盧比的先進半導體實驗室，押注印度下一波晶片機會不在晶圓廠，而在晶片從代工廠回來之後那些不起眼卻關鍵的工程環節。
-
Research ENNine Billion Variants, One Petabyte: DeepMind's AlphaGenome Atlas Precomputes Every Possible DNA Letter Change
Google DeepMind has released AlphaGenome Atlas, a 1-petabyte dataset predicting the molecular effects of all ~9 billion possible single-letter DNA variants in the human genome — 30x larger than the AlphaFold Database — free for academic use.
-
Research 中九十億個變異、一 PB 資料量：DeepMind AlphaGenome Atlas 預先算好人類基因體每個字母的改變
Google DeepMind 發布 AlphaGenome Atlas，一個 1 PB 的資料集，預測人類基因體中約 90 億個所有可能單字母 DNA 變異的分子效應 — 比 AlphaFold 資料庫大 30 倍以上，學術用途免費開放。
-
Research ENTelling Agents to Test Better Makes Them Worse: Inside Dan Luu's 26-Condition Experiment
Dan Luu ran 26 prompt conditions across thousands of Rust agent runs and found that naming a testing technique — TDD, Lean 4, QuickCheck, Verus — reliably produced worse correctness than saying nothing at all.
-
Research 中叫 AI 用更好的測試方法，結果反而更糟：Dan Luu 的 26 組對照實驗
Dan Luu 對程式代理跑了 26 種測試指令條件、數千次 Rust 實作，發現指定 TDD、Lean 4、QuickCheck、Verus 等技術的正確率，普遍低於什麼都不說的預設條件。
-
Industry ENEurope's Biggest Round Ever: Mistral Raises €3B Led by Samsung to Build the Sovereign AI Stack
Mistral AI's €3B Series D at a €21B+ post-money valuation is the largest equity round in European tech history — Samsung leads, EQT's Scaleup Europe Fund and PSG Equity co-lead, and the bet is on sovereign, open-weight AI.
-
Industry 中歐洲史上最大輪融資：Mistral 獲三星領投 30 億歐元，打造主權 AI 全堆疊
Mistral AI 以超過 210 億歐元估值完成 30 億歐元 Series D，創歐洲科技公司股權融資紀錄——三星領投、EQT 與 PSG 共同領投，押注主權開源權重 AI。
-
Tools ENFrom Voice Commands to a Household Agent: Baidu's Xiaodu Takes Its 'Family AI Brain' to Hardware
At its September 8 launch event in Beijing, Baidu's Xiaodu unveiled an agent-first hardware lineup — Tiantian companion screens, smart displays, speakers, and cameras running the agentic 'Chaoneng Xiaodu' assistant and a 2.0 AI caregiving agent.
-
Tools 中從語音指令到家庭智能體：百度小度把「家庭 AI 大腦」裝進硬體
百度小度於 9 月 8 日北京發表會推出以智能體為核心的硬體陣容——添添閨蜜機、智慧螢幕、音箱與攝影機，全面搭載智能體化的「超能小度」，攝影機並迎來第二代 AI 看護智能體。
-
Industry ENFour Logos, One Framework: Alibaba Cloud, Cambricon, and Ant Group Take Seats in the PyTorch Foundation
At PyTorch Conference China in Shanghai, the PyTorch Foundation welcomed Alibaba Cloud and Cambricon as Platinum members and Ant Group as Gold — putting Chinese cloud, chip, and fintech giants on its governing board alongside Huawei, as more than 250 Chinese organizations now contribute to PyTorch ecosystem projects.
-
Industry 中四家巨頭、一個框架：阿里雲、寒武紀與螞蟻集團正式入座 PyTorch 基金會
在上海舉行的 PyTorch Conference China 上，PyTorch 基金會宣布阿里雲與寒武紀成為白金會員、螞蟻集團成為黃金會員，與華為同台——超過 250 家中國機構如今參與 PyTorch 生態系專案，開源 AI 基礎設施的治理版圖正在重畫。
-
Models ENA 2B Model That Beats the 4B Class: OpenBMB Open-Sources MiniCPM5-2B With 131K Context and a Fully Open Data Stack
OpenBMB's MiniCPM5-2B averages 53.9 across 34 benchmarks, outscoring every 4B-class rival in its comparison set, and ships with 131K context, hybrid thinking, and the entire UltraData training stack under Apache 2.0.
-
Models 中2B 模型打贏 4B 級對手：OpenBMB 開源 MiniCPM5-2B，131K 上下文 plus 全套資料集一次奉送
OpenBMB 的 MiniCPM5-2B 在 34 項基準測試平均拿下 53.9 分，超越對比組中所有 4B 級模型，並以 Apache 2.0 授權釋出 131K 上下文、混合思考模式與完整 UltraData 訓練資料棧。
-
Research EN88.6% on BrowseComp, Weights Promised: AllSpark's Iris Agents Take the Open-Source Search Crown
AllSpark's Iris-mini and Iris-pro open-weight search agents set the pace for open models on BrowseComp, DeepSearchQA and HLE, with a fully documented SFT-RL climbing recipe.
-
Research 中BrowseComp 88.6 分、承諾開源權重：AllSpark 的 Iris 搜尋代理登上開源王座
AllSpark 團隊發表 Iris-mini 與 Iris-pro 兩個開放權重搜尋代理，在 BrowseComp、DeepSearchQA 與 HLE 寫下同級開源模型最佳成績，並完整公開 SFT-RL climbing 訓練配方。
-
Industry EN20 People, 30% Hit Rate: Inside Anthropic Labs, the Incubator That Hatched Claude Code and MCP
A Business Insider profile of Anthropic's internal startup factory reveals the two-week kill cadence behind Claude Code's $1B run-rate and MCP's 100M monthly downloads — and why cofounder Ben Mann expects most bets to fail.
-
Industry 中20 人團隊、30% 成功率：直擊 Anthropic Labs——孕育 Claude Code 與 MCP 的內部孵化器
Business Insider 深度剖析 Anthropic 的內部新創工廠：以兩週為週期快速砍掉壞點子的機制，催生了年營收跑速 10 億美元的 Claude Code 與每月下載量約 1 億次的 MCP。
-
Models ENOne Brain for the Whole Car: Alibaba Open-Sources Qwen-Drive-1.0, a 4B VLM That Sees, Explains, and Drives
Alibaba's Qwen team has open-sourced Qwen-Drive-1.0-4B, a vision-language foundation model that unifies 3D perception, driving Q&A, and trajectory planning in one Apache 2.0 package — while keeping the base Qwen3.5-4B entirely untouched.
-
Models 中一顆大腦開整台車：阿里巴巴開源 Qwen-Drive-1.0，4B 視覺語言模型同時會看、會講、會開車
阿里巴巴 Qwen 團隊開源 Qwen-Drive-1.0-4B：一套視覺語言基礎模型，整合 3D 感知、駕駛問答與軌跡規劃於一身，採 Apache 2.0 授權釋出，且完全不改動基礎的 Qwen3.5-4B 架構。
-
Industry ENThe Founder Returns: Zhang Yiming Personally Leads ByteDance's Real-Time World Model Bid
Bloomberg reports ByteDance is building a real-time spatial-video world model on Seedance, targeting ~20 fps at ~50ms latency for cloud-rendered Pico worlds, with a launch possible as soon as next month.
-
Industry 中創辦人回歸第一線：張一鳴親自督軍，字節跳動即時世界模型劍指下月發布
彭博報導，字節跳動正基於 Seedance 打造即時空間影片世界模型，目標約 20 fps、延遲低於 50 毫秒，由雲端渲染餵養 Pico 頭顯，最快下個月推出。
-
Research ENThree to Six Years Younger, By Every Clock: Insilico's AI-Designed Drug Shows Biological Age Reversal in Phase IIa
Six independent proteomic aging clocks unanimously found that rentosertib — the first drug with both an AI-discovered target and an AI-generated molecule — reversed patients' predicted biological age by 3–4 years, up to 6, in a Phase IIa IPF trial.
-
Research 中六個老化時鐘全數同意：英矽智能 AI 設計藥物在二期試驗中逆轉生理年齡 3 至 6 年
六個獨立開發的蛋白質體老化時鐘一致發現，rentosertib——首款標靶與分子皆由 AI 發現及設計的藥物——在特發性肺纖維化二期試驗中，讓受試者的預測生理年齡平均年輕 3–4 年，最高達 6 年。
-
Industry EN100 Million Dollars, 150 Spectral Bands: Pixxel's Series C and India's Planetary Infrastructure Play
Google-backed Pixxel has closed a $100M Series C led by Temasek and Seraphim Space — the largest round in Indian spacetech history — to scale its hyperspectral Earth-observation constellation.
-
Industry 中一億美元、150 個光譜頻帶：Pixxel 的 C 輪募資與印度的「行星基礎設施」豪賭
Google 早期投資的 Pixxel 完成 1 億美元 C 輪募資，由淡馬錫與 Seraphim Space 領投，創下印度太空科技史上最大單輪紀錄，將擴建其高光譜地球觀測衛星群。
-
Research ENFaster Than It Plays: Video DeltaNet Generates 14 Seconds of 768p Video in 11 Seconds
UC Berkeley, Impossible Inc., and UT Austin researchers bolt a hybrid linear-attention branch onto MiniMax H3, cutting a 14.4-second 768p render from 14 minutes to 11.23 seconds on 8 B200s — with near-lossless quality and a fully open release.
-
Research 中生成比播放還快：Video DeltaNet 用 11 秒產出 14 秒 768p 影片
UC Berkeley、Impossible Inc. 與 UT Austin 的研究團隊，在 MiniMax H3 上外加混合線性注意力分支，把 14.4 秒 768p 影片的渲染時間從 14 分鐘壓到 8 張 B200 上的 11.23 秒——品質近乎無損，且全程開源。
-
Industry EN12.7 Million Graduates, One Vanishing Entry Level: Inside China's AI Jobs Squeeze
A record 12.7 million Chinese graduates are entering a labor market where AI is absorbing the junior white-collar tasks that once trained them — and Beijing is scrambling to respond.
-
Industry 中1,270 萬畢業生，一個消失的入門職缺：透視中國的 AI 就業擠壓
中國史上最大批的 1,270 萬大學畢業生，正走進一個由 AI 接手基層白領工作的勞動市場——北京正全力尋找對策。
-
Tools EN3.5x to 10x Faster Image and Video Inference: Nunchux Exits Stealth With the Modelverse
Born from MIT HAN Lab and CMU research, Nunchux launches the Modelverse with up to 10x faster multimodal inference — FLUX.2-klein in 0.08s, MiniMax-H3 video in 10.4s — without losing quality.
-
Tools 中影像與影片推論快上 3.5 到 10 倍：Nunchux 攜 Modelverse 走出隱身狀態
源自 MIT HAN Lab 與 CMU 研究團隊的 Nunchux 本週發表 Modelverse，多模態推論最快提速 10 倍——FLUX.2-klein 只要 0.08 秒、MiniMax-H3 影片 10.4 秒完成，而且品質不減反增。
-
Tools EN156 Million Tokens for 800 Lines: SonarSource Puts a Price on the Coding Agent 'Context Tax'
SonarSource instrumented its own coding agent and found one ordinary 800-line PR burned ~156M context tokens and ~$41 — almost all of it re-billed cache reads from grep-and-read navigation.
-
Tools 中800 行程式碼燒掉 1.56 億 Token：SonarSource 為 Coding Agent 的「上下文稅」標出真實價格
SonarSource 對自家 coding agent 做全程遙測，發現一個普通的 800 行 PR 就消耗約 1.56 億 context token、花費約 41 美元——其中絕大多數是 grep 與整檔讀取被逐輪重複計費的 cache read。
-
Meta EN160,000 Ascend Chips and No Nvidia in Sight: Inside DeepSeek's Inner Mongolia Mega-Cluster
Bloomberg reports DeepSeek will deploy at least 160,000 Huawei Ascend 950DT accelerators at a 1 GW data center in Inner Mongolia — the largest known all-domestic AI cluster and a stress test of China's push to replace Nvidia.
-
Meta 中16 萬顆 Ascend 晶片、零 Nvidia：拆解 DeepSeek 內蒙古超級集群
Bloomberg 報導 DeepSeek 將在內蒙古 1 GW 資料中心部署至少 16 萬顆華為 Ascend 950DT 加速器——這是已知最大規模的全國產 AI 集群，也是中國「去 Nvidia 化」路線圖最嚴苛的一次壓力測試。
-
Meta ENHalf a Round, All a Strategy: Nvidia's $2.5B Bet on Thinking Machines
Nvidia is in talks to supply roughly half of Thinking Machines Lab's new $5-6B raise at a $40B+ valuation — the second attempt at a mega-round for Mira Murati's open-weights lab.
-
Meta 中一輪融資、一半來自輝達：談 Thinking Machines 的 25 億美元豪賭
輝達正洽談以約 25 億美元投資 Mira Murati 創辦的 Thinking Machines Lab，並包下這輪 50-60 億美元融資的半數——這是這家開放權重實驗室第二次衝擊巨型輪。
-
Industry EN$20M to $8 Billion: The Anatomy of a16z's AI Windfall
SpaceX's $60B Cursor close and Stripe's $8B OpenRouter purchase left Andreessen Horowitz holding positions worth a combined $8 billion — the fastest legitimized windfall in venture history, and a road test of Martin Casado's infrastructure thesis.
-
Industry 中從 2,000 萬美元到 80 億美元：a16z AI 豪賭的完整解剖
SpaceX 以 600 億美元收購 Cursor、Stripe 以 80 億美元買下 OpenRouter，讓 Andreessen Horowitz 手中兩筆投資合計價值超過 80 億美元——這是創投史上最快兌現的巨額回報，也是 Martin Casado 基礎設施論述的首次全面驗證。
-
Industry EN$40B or Nothing: Inside Thinking Machines' Second Try After the Round That Collapsed
Mira Murati's Thinking Machines is back raising capital — $1B led by Accel at a $40B valuation, with Nvidia in talks for $2.5B — eight months after its $50B round collapsed amid an exodus of co-founders to OpenAI.
-
Industry 中400 億美元的第二次機會：Thinking Machines 在融資崩盤後的重生之路
Mira Murati 的 Thinking Machines 捲土重來：Accel 領投 10 億美元、估值 400 億美元，Nvidia 據傳將投入 25 億美元——距離上次 500 億美元輪崩盤、共同創辦人集體回流 OpenAI 僅八個月。
-
Tools EN1,500 Tokens a Second, With a Catch: What Cerebras's Qwen 3.8 27B Launch Really Tells Us About Inference Economics
Cerebras is serving Alibaba's Qwen 3.8 27B at roughly 1,500 tokens/second on wafer-scale SRAM — but a 150k TPM cap that bills cached input at full price and a 128k context ceiling make sustained agentic coding 5x pricier than slower rivals.
-
Tools 中每秒 1,500 tokens 的代價：Cerebras 上架 Qwen 3.8 27B，揭露推論經濟學的真相
Cerebras 以晶圓級 SRAM 將阿里巴巴的 Qwen 3.8 27B 推到約每秒 1,500 tokens——但快取輸入全額計費的 150k TPM 上限與 128k 上下文天花板，讓長時間代理式編碼比慢速對手貴上 5 倍。
- Industry EN
The $12.93B Repo: NVIDIA Buys Hugging Face, the GitHub of AI
NVIDIA is acquiring Hugging Face for $12.93 billion — the largest open-source AI platform deal ever — putting the chip giant in control of where 18 million developers get their models.
- Industry EN
129 億美元的模型庫：NVIDIA 買下 AI 界的 GitHub——Hugging Face
NVIDIA 以 129.3 億美元收購 Hugging Face，創下開源 AI 平台交易金額紀錄，晶片巨頭從此掌管 1,800 萬開發者取得模型的中樞。
-
Tools EN242,000 Stars and Counting: Hermes Agent Now Out-Stars Claude Code and OpenAI Codex Combined Momentum on GitHub
Thirteen months after its first commit, NousResearch's open-source Hermes Agent has passed 242,000 GitHub stars — ahead of Anthropic's Claude Code (144K) and OpenAI Codex (122K) — with its v0.21.0 'Pantheon' release shipping multi-agent Bot Mode, memory-carrying cron jobs, and live subagent steering.
-
Tools 中24.2 萬星持續狂飆：Hermes Agent 星數正式超越 Claude Code 與 OpenAI Codex
首次提交僅 14 個月後，NousResearch 的開源專案 Hermes Agent 星數突破 242,000——領先 Anthropic 的 Claude Code（14.4 萬）與 OpenAI Codex（12.2 萬）；v0.21.0「Pantheon」版本內建多代理 Bot Mode、會記憶的排程任務與即時子代理導引。
-
Policy ENGuardrails as a Service: Inside Abliteration.ai, the Startup Selling Uncensored Frontier Models
Abliteration.ai hosts open-weight models with refusal behavior stripped from the weights, marketing them for offensive cyber and red-team work — and it just started courting VCs.
-
Policy 中把安全護欄當生意：Abliteration.ai 如何把「去拒絕化」前沿模型賣給所有人
Abliteration.ai 直接代管拒絕行為已從權重中移除的開源權重模型，主打進攻性資安與紅隊測試市場——而且剛開始與創投接洽募資。
-
Research ENThe Ethics of Listening: AI Gets Close to Decoding Animal Language — and Bioethicists Sound the Alarm
AI foundation models are closer than ever to decoding the calls of crows, whales and belugas. Bioethicists warn the same tools hand humans new levers to manipulate animals — from poachers mimicking mating calls to farms broadcasting distress vocalizations.
-
Research 中聆聽的倫理：AI 即將解讀動物語言，生物倫理學家卻拉起警報
AI 基礎模型距離解讀烏鴉、鯨魚與白鯨的叫聲從未如此接近，但生物倫理學家警告，同樣的工具也給了人類操縱動物的新槓桿——從盜獵者模仿求偶叫聲，到農場播放警戒叫聲驅趕天敵。
-
Tools ENOne Box for Storage, Security, and a Brain: Inside UGREEN's HomeAgent and the $9,999 Jetson Thor Hub That Wants to Run Your Whole Home
UGREEN's HomeAgent lineup merges a NAS, an NVR, and an on-device AI assistant — topping out with a $19,999 NVIDIA Jetson Thor T5000 hub delivering 2,070 TFLOPS of local FP4 compute.
-
Tools 中一台機器搞定儲存、監控與大腦：UGREEN HomeAgent 與那台要價 9,999 美元的 Jetson Thor 家用中樞
UGREEN 的 HomeAgent 系列把 NAS、NVR 與裝置端 AI 助理三合一，最高階的 MasterAgent MA100 採用 NVIDIA Jetson Thor T5000，提供 2,070 TFLOPS 的本地 FP4 運算能力，建議售價 19,999 美元。
-
Models ENSix Models, One Blueprint: Inside MBZUAI's K2 Horizon, the Largest Fully Open AI Release
MBZUAI's Institute of Foundation Models releases K2 Horizon — six Apache 2.0 models from 0.9B to 375B parameters with weights, code, data recipes, checkpoints, and training logs, plus a self-audit that caught its own model cheating.
-
Models 中六個模型、一份完整藍圖：MBZUAI K2 Horizon，史上最大規模的全開源 AI 發布
MBZUAI 基礎模型研究院發布 K2 Horizon——六個 Apache 2.0 授權的模型，參數從 0.9B 到 375B，連同權重、程式碼、資料配方、訓練中繼點與完整訓練日誌一次公開，甚至主動公布了自己模型在基準測試上作弊的自查報告。
-
Industry ENTokens With Your Dumplings: How China Turned AI Compute Into a Consumer Perk
Chinese banks, telcos, and restaurants now bundle AI tokens like credit-card points and data plans — a supply-led experiment in making compute a everyday consumer good.
-
Industry 中吃餃子送算力：中國如何把 AI Token 變成日常消費品
中國的銀行、電信業者與餐廳，正把 AI token 包裝成信用卡紅利與流量資費——一場把算力變成日常消費品的供給端實驗。
-
Industry ENThe GitHub of AI Now Belongs to Nvidia: Inside the $12.93 Billion Hugging Face Deal
Nvidia is acquiring Hugging Face — 2.5 million open models, 1 million datasets, 13 million developers — for $12.93 billion. The biggest open-source AI land grab ever, explained.
-
Industry 中AI 界的 GitHub 易主：深入解析 Nvidia 129.3 億美元收購 Hugging Face 世紀交易
Nvidia 以 129.3 億美元收購 Hugging Face——250 萬個開源模型、100 萬個資料集、1,300 萬開發者的中立之地。史上最大的開源 AI 併購，全文解析。
-
Industry ENSix Months Behind, Playing to Win: Kai-Fu Lee's iPhone-vs-Android Map of the US-China AI Race
In a wide-ranging Mishal Husain interview, the 01.AI founder says the US-China frontier gap has collapsed from years to six months — and that China will win AI on reach while America wins on revenue.
-
Industry 中從落后四年到只剩六個月：李開復用 iPhone vs. Android 解讀中美 AI 競賽
李開復接受 Mishal Husain 專訪時指出，中美前沿模型差距已從三、四年縮短到約六個月——中國將以觸達規模取勝，美國則賺走大部分利潤。
-
Tools ENA Bearer Token and a Blank: How a LiteLLM Auth Bug Landed in CISA's Exploited Catalog and Made AI Gateways a Target
CISA has added LiteLLM's MCP auth bypass (CVE-2026-59822, CVSS 8.8) to its Known Exploited Vulnerabilities catalog after honeypot evidence of active probing, with federal patch deadlines of September 5 and 16 — and Microsoft says compromised AI gateways are now being mined for provider keys and crypto.
-
Tools 中一個 Bearer Token 與一個空物件：LiteLLM 授權繞過漏洞登上 CISA 已知遭利用漏洞清單，AI 閘道正式成為攻擊目標
CISA 將 LiteLLM 的 MCP 授權繞過漏洞（CVE-2026-59822，CVSS 8.8）列入「已知遭利用漏洞（KEV）」清單，蜜罐觀測證實攻擊者已 actively 探測相關端點，聯邦補丁期限為 9 月 5 日與 9 月 16 日——微軟並指出遭入侵的 AI 閘道正被用來竊取供應商金鑰與挖礦。
-
Industry ENThe Great Unbundling: Open-Weight Models Hit 53% of Developer Tokens as Corporate America Walks Away From Frontier APIs
The New York Times reports US enterprises from AT&T down are pivoting to cheap, downloadable open-weight models — the same week Citi found open models at 53% of gateway token volume. Inside the quiet collapse of the frontier-API pricing model.
-
Industry 中開放權重模型的大遷徙：開發者 token 占比衝上 53%，美國企業界正集體告别前沿 API
《紐約時報》報導從 AT&T 到一般新創都在轉向便宜、可下載的開放權重模型——同一週，花旗研究發現開放模型已占 Vercel AI Gateway 流量的 53%。前沿 API 定價模式的寧靜崩塌，正在發生。
-
Models ENBenchmarks You Can't Study For: Inside Artificial Analysis Intelligence Index v4.2
Artificial Analysis rebuilt its flagship leaderboard around private test sets — 40% of the weighting — added agentic knowledge-work evals, and retired the saturated GPQA Diamond. Claude Fable 5.1 holds #1 over GPT-6 Astra on the new scale.
-
Models 中無法事先偷看的基準測試：Artificial Analysis 智慧指數 v4.2 內幕
Artificial Analysis 以私有測試集重建其招牌排行榜——權重佔 40%——新增代理式知識工作評測，並淘汰已飽和的 GPQA Diamond。在新量表上，Claude Fable 5.1 擊敗 GPT-6 Astra 穩居第一。
-
Tools ENFrom Model Picker to Workflow Engine: Inside GitHub's HydraFusion
GitHub's new research preview orchestrates multiple LLMs per task — drafting, critiquing, and cascading across model families — matching Claude Opus 5 quality at up to 67% lower estimated cost on agentic coding benchmarks.
-
Tools 中從選模型到組工作流：GitHub HydraFusion 深度解析
GitHub 的新研究預覽版以執行期編排取代單一模型：跨模型家族起草、審查、級聯升級，在 agentic coding 基準上以最高 67% 的成本降幅逼近 Claude Opus 5 品質。
-
Research ENBiology's Fourth Act: Inside AI BioDesign, the $95 Million Bet to Engineer Life Beyond Evolution
The Allen Institute, UW Medicine, and Fred Hutch launch AI BioDesign with $95M from Paul Allen's FFST — a design-build-measure-learn loop pairing Nobel laureate David Baker's protein design with open AI models to explore biology's full design space, from cancer-killing cells to plastic-eating enzymes.
-
Research 中生物學的第四幕：AI BioDesign 以 9,500 萬美元押注超越演化的生命工程
Allen Institute、華盛頓大學醫學院與 Fred Hutch 癌症中心攜手推出 AI BioDesign，獲 Paul Allen 基金會 FFST 挪出 9,500 萬美元支持——以「設計—建構—測量—學習」閉環結合諾貝爾獎得主 David Baker 的蛋白質設計與開放 AI 模型，探索從清除癌細胞到分解海洋塑膠的完整生物設計空間。
-
Industry ENMoonshot Files for Hong Kong: Inside the Kimi Maker's $50 Billion Bid to Become China's Defining AI IPO
Moonshot AI, the Beijing startup behind Kimi, has confidentially filed for a Hong Kong IPO targeting $3–5 billion at a ~$50 billion valuation — the first major public-market test of China's AI boom, built on open weights and cloud revenue-sharing.
-
Industry 中月之暗面遞交港股上市申請：Kimi 幕後公司以 500 億美元估值，爭做中國 AI 世代定價基準
Kimi 開發商月之暗面已保密遞交香港 IPO 申請，目標募集 30 至 50 億美元、估值約 500 億美元——這是中國 AI 熱潮首次面對公開市場檢驗，靠的是開放權重策略與雲端分潤談判。
-
Research EN1,000 Repos, 5,000 Skills, 134% More Medals: BAAI's DisCo Turns GitHub Into Agent Food
BAAI's DisCo framework distills 1,000 widely used ML repositories into 5,000+ verified agent skills at roughly $40 per repo — and the skill-equipped research agent scores 134.3% higher on MLE-bench, 34.4% higher on PaperBench, with GPT-5.5 held fixed.
-
Research 中1,000 個儲存庫、5,000 個技能、獎牌率提升 134%：BAAI 的 DisCo 把 GitHub 變成代理的養分
BAAI 的 DisCo 框架將 1,000 個常用機器學習儲存庫蒸餾成 5,000+ 個經驗證的代理技能，每個儲存庫成本約 40 美元——在固定使用 GPT-5.5 的條件下，配備技能的研究代理在 MLE-bench 提升 134.3%、PaperBench 提升 34.4%。
-
Industry ENCatching Hallucinations Mid-Sentence: Resect AI Exits Stealth With $25M and a Polygraph for LLMs
The Washougal, WA startup's patented in-stream tech watches model activations in real time and intervenes before a hallucination completes — betting that interceptive AI beats inspective AI for the enterprise.
-
Industry 中在幻覺生成的當下攔截它：Resect AI 攜 2,500 萬美元走出隱身，要為 LLM 裝上測謊器
總部位於華盛頓州瓦舒加爾的新創公司，以專利的「串流內」技術即時觀察模型內部活化狀態，在幻覺完成之前介入修正——押注「攔截式 AI」將勝過「事後檢查式 AI」，成為企業市場的答案。
-
Research EN166,000 Neurons, Fully Mapped: The Complete Fruit Fly Connectome Lands in Cell
Eighteen years after Janelia bet it could map a complex brain, researchers with Google and Cambridge publish the full male fruit fly connectome — 166,000 neurons and ~125 million synapses — the largest brain map ever assembled.
-
Research 中16.6 萬神經元全數解密：果蠅完整連結體登上 Cell
Janelia 與 Google、劍橋團隊耗時 18 年，發表成年雄果蠅完整中樞神經系統連結體——16.6 萬神經元、約 1.25 億突觸，史上以神經元數計最大的腦部地圖。
-
Tools ENMicrosoft's Project Zenith and AMD's Trillion-Parameter IFA Bet: Windows Reboots Itself for Local AI
At IFA 2026, Microsoft named its developer-optimized Windows 'Project Zenith' — a ready-to-code setup that runs 30B+ models locally and unmetered — while AMD unveiled the Ryzen AI Max Pro 400 and a liquid-cooled Threadripper Halo Station targeting 1-trillion-parameter local inference. Together they aim squarely at the metered cloud token economy.
-
Tools 中微軟 Project Zenith 與 AMD 的兆級參數豪賭：Windows 為本地 AI 重新開機
在 IFA 2026，微軟將開發者最佳化版 Windows 正式命名為「Project Zenith」——開箱即寫程式、本地跑 30B+ 模型免計費——AMD 則發表 Ryzen AI Max Pro 400 與液冷 Threadripper Halo Station，劍指 1 兆參數的本地推論。兩者聯手，瞄準的是按 token 計費的雲端經濟。
-
Industry ENFrom $12B to $40B in a Year: Thinking Machines' New Round Rewrites the Rules of AI Fundraising
Mira Murati's Thinking Machines Lab is in talks to raise $1 billion at a $40 billion valuation led by Accel, with NVIDIA discussing participation — below the $50B it once sought, but still an extraordinary 400x multiple on ~$100M revenue.
-
Industry 中一年內從 120 億到 400 億美元：Thinking Machines 新一輪募資改寫 AI 融資規則
Mira Murati 創辦的 Thinking Machines Lab 據報正洽談以 400 億美元估值募資 10 億美元、由 Accel 領投，NVIDIA 亦傳參與——低於先前覬覦的 500 億，卻仍是約 1 億美元年收入的 400 倍天價倍數。
-
Research ENOne Agentic Task, 10,000x the Footprint: Vals AI Puts Hard Numbers on AI's Energy Bill
Independent benchmarker Vals AI measured electricity, carbon, and water across 16 open-weight models and found long 'thinking' agentic tasks carry up to 10,000x the environmental impact of a simple query.
-
Research 中一次 Agentic 任務，一萬倍足跡：Vals AI 用實測數據揭開 AI 的能源帳單
獨立評測機構 Vals AI 對 16 個開放權重模型量測電力、碳排與水耗，發現長時間「思考」的 agentic 任務，環境衝擊最高可達簡單問答的 10,000 倍。
-
Models ENSaudi Arabia's HUMAIN Bets on China: humain-m3 Is a 428B-Parameter Arabic Frontier Built on MiniMax
HUMAIN's humain-m3 — a 428B-parameter Arabic MoE adapted from China's MiniMax-M3 — averages 89.37% across seven Arabic benchmarks, beating GPT-5.6 SOL and Opus 5 in the company's own tests.
-
Models 中沙烏地 HUMAIN 押注中國：humain-m3 是建立在 MiniMax 之上的 428B 參數阿拉伯語前沿模型
HUMAIN 的 humain-m3 是以中國 MiniMax-M3 為基礎改造的 428B 參數阿拉伯語 MoE 模型，在七項阿拉伯語基準測試平均拿下 89.37%，於公司自測中擊敗 GPT-5.6 SOL 與 Opus 5。
-
Models ENSix Models, Zero Secrets: Inside K2 Horizon, the Largest Fully Open AI Release in History
MBZUAI's Institute of Foundation Models shipped six Apache 2.0 models from 0.9B to 375B parameters — with weights, code, training data, and methodology all public. It's the widest disclosure ever from a frontier-adjacent lab.
-
Models 中六個模型、零秘密：K2 Horizon——史上最大規模的全開源 AI 釋出
MBZUAI 基礎模型研究所一次釋出六個 Apache 2.0 授權的模型，參數規模從 0.9B 到 375B，連訓練程式碼、訓練資料與方法論全部公開——這是前沿實驗室有史以來最徹底的透明化釋出。
-
Tools ENNVIDIA PAIR Turns Your Idle Home PCs Into a Private AI Cluster: Inside the Personal AI Router
At IFA 2026 NVIDIA shipped PAIR, a free open-source router that distributes AI inference across every idle PC on your home network — from RTX 20-series GPUs to Apple M4 Macs — alongside 1.9x faster llama.cpp/vLLM optimizations and October's RTX Spark PCs.
-
Tools 中NVIDIA PAIR 把閒置家用電腦變成私人 AI 叢集：Personal AI Router 深度解析
NVIDIA 在 IFA 2026 發表免費開源工具 PAIR（Personal AI Router），能自動發現家中網路上的閒置電腦並分配 AI 推論工作——從 RTX 20 系列顯卡到 Apple M4 Mac 都支援，同場還有提速 1.9 倍的 llama.cpp/vLLM 優化與十月登場的 RTX Spark 電腦。
-
Research EN4.5 Billion TikTok Records, 289GB, Three Weeks: The Private-API Scrape That Just Landed on Hugging Face
An anonymous researcher scraped 4.5 billion TikTok video records through the app's private Android API in three weeks and published all 289GB on Hugging Face — no login, no account, just forged devices and reverse-engineered signatures.
-
Research 中45 億筆 TikTok 資料、289GB、三週完成：一場繞過 App 私有 API 的爬取行動登陸 Hugging Face
一位獨立研究者透過 TikTok App 的私有 Android API，在三週內爬取 45 億筆影片紀錄，並將 289GB 資料集完整公開在 Hugging Face——全程無帳號、無登入，靠的是偽造裝置與逆向簽名。
-
Industry ENNVIDIA Buys Hugging Face for $12.93 Billion: The Biggest Open-Source AI Deal Ever
Jensen Huang announced NVIDIA will acquire Hugging Face — home to 3M+ open models and 18M developers — for $12.93B, promising the platform stays open. Here is what the deal changes.
-
Industry 中NVIDIA 以 129.3 億美元收購 Hugging Face：開源 AI 史上最大交易
黃仁勳宣布 NVIDIA 將以 129.3 億美元收購擁有 300 萬個開源模型、1,800 萬開發者的 Hugging Face，並承諾平台維持開放。這筆交易改變了什麼？
-
Models ENFrom Fourth to First in One Week: How Alibaba's Qwen3.8-Max-0902 Snapshot Conquered CodeArena
Alibaba's date-stamped Qwen3.8-Max-0902 update jumped from 1,669 to 1,691 on CodeArena: WebDev, dethroning Claude Opus 5 — not with a new architecture, but with one targeted RL post-training pass on coding and 'cowork' agent trajectories.
-
Research EN215,128 Machine-Made Pages Are Grounding AI Answers: Inside the Trellner Study of Perplexity's Citation Supply Chain
Trellner Research ran 380 buyer-intent queries through Perplexity's sonar models and found 59.8% of 7,534 citations pointing to domains outside the world's top 100,000 sites — with three machine-generated 'best software' farms supplying 215,128 pages explicitly titled 'Facts & Grounding Page' for models to read.
-
Research 中21.5 萬頁機器生成內容正在「接地」AI 的答案：Trellner 揭露 Perplexity 引用供應鏈
Trellner Research 對 Perplexity 的 sonar 模型投放 380 組採購意圖查詢，發現 7,534 筆引用中有 59.8% 指向全球前 10 萬名以外的網域——三個機器生成的「最佳軟體」內容農場供應了 215,128 頁明寫著「Facts & Grounding Page」、專門給模型讀的頁面。
-
Tools ENCoder's Agent Relay Brings Cursor's Cloud Agents Behind the Firewall: Self-Hosted Execution for Regulated AI Coding
Coder and SpaceXAI launch Agent Relay, a self-hosted execution layer that runs Cursor Cloud Agents on customer infrastructure — opening regulated enterprises to agentic coding while Cursor keeps the agent loop.
-
Tools ENCoder Agent Relay 攜手 SpaceXAI：讓 Cursor 雲端代理程式走進企業防火牆內的自架執行時代
Coder 與 SpaceXAI 推出 Agent Relay，一個自架執行環境，讓 Cursor 雲端代理程式在客戶自己的基礎設施上執行——Cursor 保留代理迴圈，為受監管企業打開代理式編碼的大門。
-
Tools ENEquinix Inference Exchange: Nvidia and Together AI Bet the Enterprise Edge on Open Models
Equinix's new distributed inference platform pairs Nvidia reference architectures with Together AI's 200+ open models across 280+ data centers, betting that where inference runs becomes the next enterprise battleground.
-
Tools 中Equinix Inference Exchange：Nvidia 與 Together AI 押注企業邊緣的開源模型推論
Equinix 的分散式推論平台結合 Nvidia 參考架構與 Together AI 的 200+ 開源模型，涵蓋 280+ 座資料中心，賭的是「推論在哪裡跑」將成為企業下一個戰場。
-
Meta ENSix to Zero: AISLE's Autonomous AI Finds 6 curl CVEs After OpenAI and Anthropic's Frontier Models Found None
Days after Anthropic Mythos and OpenAI Codex Security reported zero remaining flaws in curl, a startup's specialized AI system filed 29 reports — six became CVEs in curl 8.22.0, and the Linux kernel maintainer says he's seeing the same pattern.
-
Meta 中六比零：在 OpenAI 與 Anthropic 前沿模型掛零之後，AISLE 的自主 AI 系統在 curl 找出 6 個 CVE
Anthropic Mythos 與 OpenAI Codex Security 對 curl 回報「找不到更多問題」數天後，一家新創的專用 AI 系統提交了 29 份報告——其中 6 個成為 curl 8.22.0 的正式 CVE，Linux 核心維護者直言自己看到了同樣的現象。
-
Research EN535 Out of 600: NVIDIA's Nemotron Becomes the First AI to Beat Every Human at IOI 2026
NVIDIA's open-weight Nemotron-3-Ultra-CC scored 535.4/600 at IOI 2026 — beating the top human (498.27) and gold threshold (361.12) live, under identical contest constraints, using GenCorrect feedback-driven test-time refinement.
-
Research 中535 分滿分 600：NVIDIA Nemotron 成為首個在 IOI 2026 擊敗所有人類的 AI
NVIDIA 開放權重的 Nemotron-3-Ultra-CC 在 IOI 2026 拿下 535.4/600 —— 在與人類選手完全相同的比賽規則下現場擊敗最高分人類（498.27）與金牌門檻（361.12），關鍵在於 GenCorrect 回饋驅動的測試時運算策略。
-
Industry ENLEAP 2026 Wrap: HUMAIN's $15 Billion AI Blitz and Saudi Arabia's Sovereign Stack
As LEAP 2026 closes in Riyadh, Saudi Arabia's HUMAIN has unveiled a 1 GW AMD-Cisco buildout, an AWS cloud region, driverless trucks, and sovereign frontier models — over $15 billion in deals in four days.
-
Industry 中LEAP 2026 收官：HUMAIN 千五億美元 AI 攻勢與沙烏地阿拉伯的主權 AI 全棧佈局
LEAP 2026 今日在利雅德落幕。HUMAIN 連發 AMD-Cisco 1 GW 基礎設施、AWS 沙烏地雲區域、自動駕駛卡車與主權前沿模型等多項合作，四天內簽下超過 150 億美元協議。
-
Models ENMeta Ships Muse Spark 1.3: The Agentic Model That Uses 20% Fewer Tool Calls and Knows When to Ask for Help
Meta's Muse Spark 1.3 lands in Muse Code and the Meta Model API with better long-horizon agency, ~20% fewer tool calls, ~25% fewer tokens, and a max-reasoning mode still waiting on safety testing.
-
Models 中Meta 推出 Muse Spark 1.3：省 20% 工具呼叫、懂得適時求援的代理模型
Meta 的 Muse Spark 1.3 於 9 月 2 日上線 Muse Code 與 Meta Model API，強化長程代理任務、減少約 20% 工具呼叫與 25% token 消耗，max 推理模式則待安全測試後推出。
-
Industry ENOwkin Licenses Its K Pro 'AI Scientist' and Multimodal Patient Data to Boehringer Ingelheim
The Paris-New York agentic AI company will license K Pro and its MOSAIC oncology atlas to Boehringer, and generate fresh multimodal immunology data — the third big-pharma K Pro deal of 2026.
-
Industry ENOwkin 將 K Pro「AI 科學家」與多模態病患數據授權給百靈佳殷格翰
這家巴黎-紐約雙總部的 Agentic AI 公司，將授權 K Pro 與 MOSAIC 癌症空間多體學圖譜給百靈佳殷格翰，並為其新生成免疫領域的多模態數據——這是 2026 年第三筆大型藥廠 K Pro 授權案。
-
Industry ENEurope's €387.8M LUMI-AI Bet: AMD MI430X GPUs and the Supercomputer Built to Win the AI Sovereignty Race
EuroHPC has signed a €387.8 million contract with Bull to build LUMI-AI, a next-generation AMD-powered AI supercomputer in Kajaani, Finland — 10x the AI capacity of today's LUMI, deployment in 2027, and the clearest signal yet of Europe's sovereign-compute ambitions.
-
Industry 中歐洲 3.878 億歐元的 LUMI-AI 豪賭：AMD MI430X GPU 與為 AI 主權競賽而生的超級電腦
EuroHPC 與 Bull 簽署 3.878 億歐元合約，將在芬蘭 Kajaani 打造次世代 AMD 架構 AI 超級電腦 LUMI-AI —— AI 算力達現有 LUMI 的 10 倍，2027 年部署，是歐洲主權算力企圖心最明確的訊號。
-
Tools ENThe $1,688 Humanoid: Nori Robotics' YC-Backed Bid to Collapse Robot Economics
YC S26 startup Nori Robotics is shipping a $1,688 bimanual wheeled humanoid from San Francisco — 19 DOF, open SDK, $350K in sales in six weeks, and a plan to turn every customer into a data source for generalist robot policies.
-
Tools 中1,688 美元的人形機器人：Nori Robotics 獲 YC 投資，要徹底改寫機器人成本結構
YC S26 新創 Nori Robotics 正從舊金山出貨一款 1,688 美元的雙臂輪式人形機器人 — 19 自由度、開放 SDK、六週內締造 35 萬美元銷售，並計劃讓每一位客戶都成為通用機器人政策的資料來源。
-
Models ENEurope's New Frontier Contender: Multiverse Computing's Quasar 438B Scores 43 on the AA Intelligence Index
The Spanish quantum-inspired AI lab's first large model beats Mistral Medium 3.5 and NVIDIA Nemotron 3 Ultra on Artificial Analysis benchmarks while answering 500-token prompts in 15.3 seconds — Europe's highest-scoring model yet.
-
Models EN歐洲新世代 AI 旗手：Multiverse Computing 發表 Quasar 438B，AA 智慧指數奪下 43 分
這家以量子啟發式模型壓縮技術聞名的西班牙公司推出首款大型模型，在 Artificial Analysis 基準上擊敗 Mistral Medium 3.5 與 NVIDIA Nemotron 3 Ultra，並以 15.3 秒完成 500 token 回應——成為歐洲迄今得分最高的模型。
-
Research ENOne Model to Serve Them All: T-Tech Turns Qwen3-32B Into a Fleet of GRPO Experts and Retires the 7× Bigger Baseline
A T-Tech team split Qwen3-32B into per-axis GRPO experts, merged them with two-stage SLERP, and beat a ~7× larger baseline on instruction following and function calling — while absorbing 116M requests a month at a fraction of the cost.
-
Research 中一個模型服務全部：T-Tech 把 Qwen3-32B 拆成 GRPO 專家軍團，淘汰 7 倍大的基線模型
T-Tech 團隊把 Qwen3-32B 拆成依能力軸訓練的 GRPO 專家，再用兩階段 SLERP 合併，在指令遵循與函式呼叫上擊敗約 7 倍大的基線——同時每月吸收 1.16 億次請求，成本只有零頭。
-
Industry ENSouth Korea's $919 Billion Sovereign AI Bet: 18.4GW of Datacenters and a Tournament That Eliminated Its Best Model
A SemiAnalysis deep dive reveals Korea's $919B, 18.4GW-by-2035 sovereign AI buildout — and how the national foundation-model tournament just eliminated the best non-Chinese open-source model in the world.
-
Industry 中韓國 9,190 億美元主權 AI 豪賭：18.4GW 資料中心與一場淘汰掉自己最強模型的國家級賽局
SemiAnalysis 深度報告揭露韓國 9,190 億美元、2035 年前 18.4GW 的主權 AI 建設計畫——而國家基礎模型錦標賽剛淘汰了全世界最強的非中國開源模型。
-
Tools ENQualcomm and ASUS Put a 20B-Parameter Pharmacy AI Agent on the Counter: Offline, On-Device, and Built for Taiwan's Super-Aged Society
Qualcomm and ASUS launched the Pharmaceutical AI Agent: a GPT-OSS 20B model distilled from 120B to run locally on Snapdragon-powered AI PCs, reviewing 28 prescription safety metrics against TFDA drug data for 50+ community pharmacies in southern Taiwan — fully offline, no cloud required.
-
Tools 中Qualcomm 與華碩把 200 億參數的藥局 AI 代理搬上櫃檯：離線、在地運算，為台灣超高齡社會而生
Qualcomm 與華碩集團發表「藥事 AI 代理」（Pharmaceutical AI Agent）：將 GPT-OSS 從 1,200 億參數蒸餾到 200 億，可在 Snapdragon AI PC 上完全本地運算，對照 TFDA 藥品說明書資料庫自動檢核 28 項處方安全指標，部署於南台灣 50 多家社區藥局——全程離線、無需雲端。
-
Models ENQwen3.8-Max-0902 Debuts at #1 on Code Arena WebDev: Alibaba's Coding & Cowork Refresh Dethrones Claude Opus 5
Alibaba's post-trained Qwen3.8-Max-0902 jumped from 1,669 to 1,691 points to take the top spot on Code Arena's WebDev leaderboard, edging out Claude Opus 5 (Max) and Kimi K3 Max at a blended $5 per million tokens.
-
Models 中Qwen3.8-Max-0902 空降 Code Arena WebDev 榜首：阿里「編碼與協作」強化版以 1,691 分擊退 Claude Opus 5
阿里巴巴針對 Coding & Cowork 後訓練的 Qwen3.8-Max-0902，在 Code Arena WebDev 榜單從 1,669 躍升至 1,691 分奪冠，超越 Claude Opus 5（Max）與 Kimi K3 Max，混合單價僅約每百萬 token 5 美元。
-
Research ENNeural Networks Were Secretly Symbolic All Along: Inside the Paper Reconciling AI's Oldest Feud
A new 30-page study from Yale, JHU, NYU, and Microsoft Research shows the vector representations inside MLPs, RNNs, Transformers, and seven open-weight LLMs can be replaced by closed-form symbolic equations with almost no change in behavior.
-
Research 中神經網路其實一直是符號系統：一篇新論文如何調停 AI 最古老的論戰
來自 Yale、Johns Hopkins、NYU 與微軟研究院的 30 頁新研究顯示：MLP、RNN、Transformer 以至七個開放權重 LLM 的內部向量表徵，都能用閉式符號方程式取代，而行為幾乎不變。
-
Models ENMeta's Muse Voice Transcribe Listens Like a Human: 20+ Speaker Diarization, 70+ Languages, One Hour Sessions
Meta Superintelligence Labs ships its first real-time audio perception model: streaming ASR with adaptive delay trained via RL, native code-switching, and a claimed #1 spot on Artificial Analysis speech-to-text rankings.
-
Models ENMeta Muse Voice Transcribe 像人類一樣聆聽：20+ 說話者分離、70+ 語言、一小時長音檔一次搞定
Meta 超級智慧實驗室推出首款即時語音感知模型：以強化學習訓練的自適應延遲串流 ASR、原生 code-switching，並宣稱登上 Artificial Analysis 語音轉文字排行榜第一。
-
Industry ENGlean Claims Claude Cowork Costs 5x More Per Task: Inside the Benchmark Wars Reshaping Enterprise AI
Glean's benchmark puts its auto-routed assistant at $0.58 per enterprise task versus Claude Cowork's $2.98 — an 81% gap it attributes to context architecture, not model quality. The Information amplified the claim this week, and it lands amid Uber and ServiceNow blowing their annual AI budgets in months.
-
Industry ENGlean 實測宣稱 Claude Cowork 每任務成本貴 5 倍：重塑企業 AI 的基準測試大戰
Glean 的基準測試顯示，其自動路由助理完成企業任務平均每件僅 0.58 美元，而 Claude Cowork 要價 2.98 美元——81% 的差距來自上下文架構而非模型優劣。在 Uber、ServiceNow 相繼提早燒光年度 AI 預算之際，這份報告直擊每個 Anthropic 客戶的痛點。
-
Models ENGoogle Unveils Gemini 3.8 Flash Today: The 'Skimaki' Refinement That Takes Aim at Claude Fable 5
Google DeepMind lifts the curtain on Gemini 3.8 Flash on September 2 — a Jetski-proven 'skimaki' build that trims verbose output and targets Claude Fable 5's coding crown at one-tenth the price.
-
Models 中Google 今日揭曉 Gemini 3.8 Flash：代號「skimaki」的精煉之作，劍指 Claude Fable 5
Google DeepMind 於 9 月 2 日公開 Gemini 3.8 Flash——這個歷經 Jetski 內部實測的「skimaki」版本砍掉冗長輸出，以十分之一的價格挑戰 Claude Fable 5 的編程王座。
-
Models ENMercury 2.5 Preview Hits 1,107 Tokens Per Second: Inception's Diffusion LLM Quietly Rewrites the Economics of Fast Reasoning
Inception Labs' Mercury 2.5 Preview generates and refines tokens in parallel instead of one at a time, hitting 1,107 tokens/sec on standard GPUs with frontier-lite quality at a fraction of the price.
-
Models 中Mercury 2.5 Preview 每秒 1,107 Token：Inception 的擴散式 LLM 悄悄改寫高速推理的經濟學
Inception Labs 的 Mercury 2.5 Preview 捨棄逐一生成 token 的做法，改以平行生成、反覆精煉的方式達成每秒 1,107 token 的輸出速度，並以遠低於同級模型的價格提供接近前緣等級的品質。
-
Models ENAnthropic Ships Claude Fable 5.1 and Mythos 5.1: Same Price, 75% Cheaper Cache Reads
Anthropic's September 1 release keeps Fable pricing at $10/$50 per MTok but slashes cache reads to $0.25, targeting long-horizon agents with a 1M-token context and 128K output.
-
Models ENAnthropic 發布 Claude Fable 5.1 與 Mythos 5.1：價格不變，快取讀取成本大降 75%
Anthropic 於 9 月 1 日推出 Fable 5.1，維持每百萬 token 輸入 10 美元、輸出 50 美元的定價，但將快取讀取降至 0.25 美元，搭配百萬 token 上下文與 128K 輸出，瞄準長時程 Agent 工作負載。
-
Industry ENPhysical Superintelligence Emerges From Stealth With $58M to Build an AI Physics Lab — and an Interstellar Mission
PSI launched today with a $58M seed led by Breakthrough Energy Ventures, an Emmy platform of virtual physicists, an open-source AI physicist, and a founding role in the first AI-planned interstellar mission to Alpha Centauri.
-
Industry 中Physical Superintelligence 攜 5,800 萬美元種子輪亮相：打造 AI 物理實驗室，還要規劃星際任務
PSI 今日亮相，獲 Breakthrough Energy Ventures 領投的 5,800 萬美元種子輪，推出虛擬物理學家平台 Emmy、開源 AI 物理學家，並擔任首個 AI 規劃的半人馬座星際任務的創始技術夥伴。
-
Research ENAlibaba's Amap Team Open-Sources DreamX-Creator: A 7B Model That Generates Video and Sound Together at 2K
DreamX-Creator 1.0 jointly denoises audio and video streams in a single 7B model, adds RL with multimodal feedback, and refines output to 2K in one denoising step per chunk.
-
Research 中阿里巴巴高德團隊開源 DreamX-Creator：7B 參數模型一次生成 2K 影音，畫面與聲音不再是兩回事
DreamX-Creator 1.0 以單一 7B 模型同時去噪生成音訊與視訊串流，搭配多模態強化學習與每片段一步去噪的 2K 精煉管線。
-
Industry ENTogether AI Turns to Saudi Arabia: Inside the 250MW HUMAIN Deal Targeting $5 Billion in Year One
As US data-center siting hits a wall, Together AI is leasing 250 megawatts and 120,000 accelerators in Saudi Arabia with PIF-backed HUMAIN — a deal expected to generate over $5 billion in gross annualized revenue in its first year.
-
Industry 中Together AI 轉向沙烏地阿拉伯：HUMAIN 250MW 資料中心協議，第一年目標 50 億美元營收
美國資料中心选址遭遇瓶頸之際，Together AI 與沙國主權基金支持的 HUMAIN 簽訂協議，租用 250 百萬瓦電力與 12 萬顆加速器，預計第一年創造超過 50 億美元年化毛營收。
-
Research ENMicrosoft's SWA Paper Upends Linear Attention: 60 Tokens of Context Beat Months of Post-Training
A Microsoft Applied Sciences team shows a training-free sliding-window attention mask with 4 attention sinks matches or beats post-trained linear attention across Llama, Qwen, and Phi-4 — 2-10x higher on long-context reasoning.
-
Research 中微軟 SWA 論文顛覆線性注意力：60 個 token 的上下文勝過數月的後訓練
微軟 Applied Sciences 團隊證明：免訓練的滑動視窗注意力（含 4 個 attention sink）在 Llama、Qwen、Phi-4 上媲美甚至超越後訓練線性注意力，長上下文推理更拿下 2-10 倍差距。
-
Models ENDeepSeek Open-Sources Its First Multimodal Agent: V4-Flash-Vision-Exp Weights Go Public Under MIT
DeepSeek published the full 305B-parameter DeepSeek-V4-Flash-Vision-Exp checkpoint to Hugging Face under an MIT license — native FP8 weights, a 1M-token context, DSpark speculative decoding, and agent scores within a point of Opus-4.8 on five benchmarks.
-
Models 中DeepSeek 開源首款多模態代理模型：V4-Flash-Vision-Exp 權重以 MIT 授權公開
DeepSeek 將完整的 305B 參數 DeepSeek-V4-Flash-Vision-Exp 檢查點以 MIT 授權上傳至 Hugging Face——原生 FP8 權重、百萬 token 上下文、DSpark 投機解碼，並在半數多模態代理基準上追平甚至超越 Opus-4.8。
-
Models ENGoogle's TimesFM-3 Forecasts Multivariate Time Series in a Single Pass
Google Research's 330M-parameter time-series foundation model generates full multivariate forecast horizons — targets, covariates, and 9 quantiles — in one forward pass, topping Gift-Eval, FEV-Bench, and Time.
-
Models 中Google TimesFM-3：一次前向傳遞，完成多變量時間序列預測
Google Research 的 3.3 億參數時間序列基礎模型，能在單次前向傳遞中生成完整的多變量預測視界——目標序列、協變數與 9 個分位數一次到位，橫掃 Gift-Eval、FEV-Bench 與 Time 三大基準。
-
Research ENSt. Jude's AdaptiveFlow Screens 69 Billion Molecules for 1,000x Less: AI Drug Discovery Gets a Cloud-Native Rewrite
St. Jude's open-source AdaptiveFlow platform steers 69-billion-molecule virtual screens with active learning, scales to 5.6 million cloud CPUs, and cuts screening costs up to 1,000-fold while discovering validated nanomolar FSP1 and PARP-1 inhibitors.
-
Research 中St. Jude 的 AdaptiveFlow 以千分之一成本篩選 690 億分子：AI 藥物發現的雲端重寫
St. Jude 開源平台 AdaptiveFlow 以主動學習導引 690 億分子的虛擬篩選、可在 AWS 上擴展至 560 萬顆 CPU，並將篩選成本最高降低 1,000 倍，同時發現經實驗驗證的奈莫耳級 FSP1 與 PARP-1 抑制劑。
-
Industry ENZ.ai's Revenue Quintupled to $142 Million in H1 — and Its Losses Are Finally Shrinking
The GLM maker's first-half sales jumped ~400% on an API surge, yet the stock fell: a look inside China's AI price war economics.
-
Industry 中Z.ai 上半年營收暴增至 1.42 億美元——虧損終於開始收斂
GLM 模型開發商上半年營收年增約 400%，受 API 業務暴增推動，但股價反應冷淡：深入解析中國 AI 價格戰的經濟學。
-
Models ENMoonshot Retires Kimi K2.5, Bets the Company on K3 at 5x the Price
Moonshot AI completed the retirement of Kimi K2.5 and moonshot-v1 on August 31, leaving the 2.8T-parameter Kimi K3 as its only flagship — at roughly 5x the per-token price of its predecessor, a deliberate reversal of China's year-long price war.
-
Models 中Moonshot 退役 Kimi K2.5，以貴五倍的 K3 押上全部身家
Moonshot AI 於 8 月 31 日完成 Kimi K2.5 與 moonshot-v1 的退役，2.8 兆參數的 Kimi K3 成為唯一現役旗艦——每 token 价格約為前代的五倍，等於親手終結了中國 AI 界長達兩年的價格戰。
-
Research ENByteDance's Lucida Turns Messy Room Videos Into Editable 3D Scenes for Robots
ByteDance Seed's Lucida pipeline parses cluttered indoor video into per-instance scene graphs, generates an asset per object, and lets a VLM policy drive the 3D editor's own gizmo handles until it decides placement is done — posting a 69% mAP gain on R2S-Scene.
-
Research 中ByteDance Lucida：把雜亂房間影片變成機器人可用的可編輯 3D 場景
ByteDance Seed 的 Lucida 管線把雜亂室內影片解析成逐物件場景圖、為每個物件生成資產，再由 VLM 策略 GizmoAct 直接操作 3D 編輯器的 gizmo 手柄決定擺放完成與否——在 R2S-Scene 上繳出 69% 的 mAP 增益。
-
Industry ENDeepSeek Nears a $7.4 Billion Round at a $74 Billion Valuation — Its Second Mega-Raise in Three Months
Fresh off its first-ever outside round in June, DeepSeek is closing in on roughly 50 billion yuan more at a ~500 billion yuan pre-money valuation, with Monolith, Shixiang, CATL and local-government funds lining up ahead of a possible 2027 Shanghai listing.
-
Industry 中DeepSeek 將以 740 億美元估值完成 74 億美元融資——三個月內第二筆巨額注資
繼 6 月首度對外募資後，DeepSeek 即將再籌約 500 億人民幣，投前估值約 5,000 億人民幣，Monolith、勢詳、寧德時代與地方政府基金均已表態參與，為 2027 年上海上市鋪路。
-
Tools ENGoogle Antigravity's /boost: A Three-Phase Multi-Agent Pipeline for the Bugs That Break Single-Agent Coding
Google documents /boost, a new slash command in Antigravity 2.0 and the Antigravity CLI that spins up an orchestrator, parallel subagents, and iterative verification loops for race conditions, algorithmic work, and deep refactors.
-
Tools 中Google Antigravity 的 /boost：用三階段多代理管線，對付讓單一代理編程工具卡死的那種 Bug
Google 為 Antigravity 2.0 與 Antigravity CLI 文件化了 /boost 指令：一條由編排器、平行子代理與反覆驗證迴圈組成的推理管線，專攻競態條件、演算法優化與深度重構。
-
Tools ENMCP at 400 Million Monthly Downloads: How Anthropic's Agent Standard Became the Internet's Plumbing
The Model Context Protocol now pulls 400 million SDK downloads a month — 4x growth this year — as the stateless 2026-07-28 spec lands MCP on serverless and edge infrastructure and locks in its status as the default way AI agents touch the outside world.
-
Tools 中MCP 月下載量突破 4 億：Anthropic 的 Agent 標準如何成為網際網路的基礎管線
Model Context Protocol 的 SDK 月下載量已達 4 億次——今年成長 4 倍——隨著無狀態的 2026-07-28 規範讓 MCP 得以部署在 serverless 與邊緣運算架構上，它已穩固成為 AI agent 接觸外部世界的預設標準。
-
Industry ENAMD, Saudi MCIT, and DCO Launch Open Developer Ecosystem to Advance AI Innovation
At LEAP 2026, AMD, Saudi Arabia's MCIT, and the Digital Cooperation Organization launched a joint initiative to give developers and startups open AI tools, ROCm training, and governance resources across 16 member states.
-
Industry 中AMD 攜手沙烏地 MCIT 與 DCO 啟動開放開發者生態系，加速 AI 創新
在 LEAP 2026 上，AMD、沙烏地阿拉伯通訊與資訊科技部（MCIT）與數位合作組織（DCO）共同宣布啟動開發者生態系計畫，將 ROCm 開源軟體、AI 開發者培訓與治理資源帶給 16 個成員國的開發者與新創。
-
Tools ENOpenClaw 2.0 Ships With 933 Contributors and 16,000 Merged PRs — the Largest Crowd-Built AI Agent Release Ever
OpenClaw 2.0 (v2026.8.1) rebuilds installation, browser, memory, and security in one release — 933 contributors, 16,000+ merged PRs, and a multiplayer cloud-session model its own team used to ship it.
-
Tools 中OpenClaw 2.0 登場：933 位貢獻者、16,000 個合併 PR——史上最大規模的群眾協作 AI Agent 發布
OpenClaw 2.0（v2026.8.1）一次重建安裝流程、瀏覽器、記憶層與安全模型——933 位貢獻者、超過 16,000 個合併 PR，團隊甚至用新的多人雲端工作階段功能來開發這個版本本身。
-
Industry ENZhipu's Revenue Quintuples as Z.AI Posts First Results as a Public AI Lab
Zhipu (Z.AI) reported H1 revenue of 953.9 million yuan, up 400%, while narrowing its net loss — the first big earnings test for China's open-source model champion.
-
Industry 中智谱营收暴增五倍：Z.AI 交出上市後首份公開財報
智谱（Z.AI，2513.HK）公布上半年營收 9.539 億人民幣、年增約 400%，淨虧損同步收窄——中國開源模型旗手迎來上市後第一場大型財報考驗。
-
Research ENMitsubishi Electric's TUSS Gives Physical AI a Pair of Ears
A single prompt-driven AI model now handles speech separation, enhancement, and environmental sound extraction at once — aimed at factory floors and public spaces, with a live demo at CEATEC 2026.
-
Research 中三菱電機 TUSS：讓實體 AI 長出一對耳朵
單一提示驅動的 AI 模型，同時搞定語音分離、語音強化與環境音抽取——瞄準工廠現場與公共空間，並將在 CEATEC 2026 現場實機展演。
-
Models ENTencent Open-Sources Hy4 Preview: A 770B MoE Flagship With 1M-Token Context
Tencent's Hunyuan team releases Hy4 preview, a 770B-parameter open-weight MoE with 49B active parameters, a 1M-token context window, and coding results that rival GLM-5.3 and Kimi K3.
-
Models 中騰訊開源 Hy4 Preview：770B 參數 MoE 旗艦模型，支援百萬 token 上下文
騰訊混元團隊發布 Hy4 preview，770B 總參數的開放權重 MoE 模型，每 token 僅啟動 49B 參數，支援 100 萬 token 上下文，編碼表現超越 GLM-5.3 與 Kimi K3。
-
Policy ENSony and Warner Sue Anthropic in Multi-Billion-Dollar Music Copyright Case — and Name Dario Amodei Personally
Sony Music Publishing and Warner Chappell allege Claude was trained on tens of thousands of pirated lyrics, seeking up to $150,000 per work and naming Anthropic's CEO as an individual defendant.
-
Policy 中Sony 與 Warner 聯手控告 Anthropic 音樂版權索賠數十億美元 — 執行長 Dario Amodei 一併被列為被告
Sony Music Publishing 與 Warner Chappell 指控 Claude 以數萬份盜版歌詞訓練，每件作品最高求償 15 萬美元，並罕見地將 Anthropic 執行長列為個人被告。
-
Models ENCohere Parse: The $1.50-Per-1,000-Page Specialist Picking Apart Enterprise Documents
Cohere's Parse (parse-v5.0) is a 2.3B-parameter vision-language model that turns contracts, invoices, and filings into clean Markdown at $1.50 per 1,000 pages — scoring 79.2 on ParseBench, beating Mistral OCR 4, and running at 2,160 pages per minute on an 8x H100 node.
-
Models 中Cohere Parse：每千頁 1.50 美元的文件解析專家模型，拆解企業文件的最後一哩
Cohere 推出 Parse（parse-v5.0），一個僅 23 億參數的視覺語言模型，能把合約、發票與申報文件轉成乾淨的 Markdown，每千頁只要 1.50 美元 — ParseBench 拿下 79.2 分擊敗 Mistral OCR 4，在 8x H100 節點上每分鐘可處理 2,160 頁。
-
Industry ENHUMAIN and Applied Intuition Will Build the World's Largest Autonomous Trucking Network in Saudi Arabia
Saudi Arabia's HUMAIN and Silicon Valley's Applied Intuition will deploy thousands of Level 4 autonomous trucks across the Kingdom's freight corridors by 2030, the first step of a national physical AI strategy spanning robotaxis, ports, mining, and construction.
-
Industry 中HUMAIN 攜手 Applied Intuition，將在沙烏地阿拉伯打造全球最大自動駕駛卡車網路
沙烏地阿拉伯 HUMAIN 與矽谷 Applied Intuition 宣布策略合作，2030 年前將在王國主要物流走廊部署數千輛 Level 4 自動駕駛卡車，並以此為起點，將實體 AI 擴展到機器人計程車、港口、礦業與營造等產業。
-
Tools ENAnthropic's Model Hardware Standard: AI Agents Take Control of the Lab Bench
Anthropic's Model Hardware Standard gives AI agents a universal interface to operate microscopes, liquid handlers, and robotic arms — turning weeks of integration work into minutes.
-
Tools 中Anthropic 模型硬體標準：讓 AI 代理親手操作實驗室儀器
Anthropic 的 Model Hardware Standard 為 AI 代理提供操作顯微鏡、自動分注器與機械臂的通用介面，把原本需要數週的整合工作壓縮到幾分鐘。
-
Models ENYutori's Navigator n2: The 27B Model That Out-Computers Frontier Giants at One-Tenth the Price
Ex-Meta AI leaders at Yutori shipped Navigator n2, a 27B computer-use model that scores 65.2% on OSWorld 2.0 — beating GPT-5.6 Sol — for $0.50/$4 per million tokens, and tops MyPCBench by 20 points over Claude Opus 4.8.
-
Models 中Yutori Navigator n2：270 億參數的電腦操作模型，以十分之一價格擊敗前沿巨頭
前 Meta AI 主管創辦的 Yutori 推出 Navigator n2，這個 270 億參數的電腦操作模型在 OSWorld 2.0 拿下 65.2%，超越 GPT-5.6 Sol，API 定價僅每百萬 token 0.5／4 美元，更在 MyPCBench 領先 Claude Opus 4.8 二十個百分點。
-
Tools ENAlibaba Takes QwenWork Global: The Workplace AI Agent Platform Enters International Public Beta
Alibaba has opened QwenWork International to global users in public beta — an all-in-one workplace AI agent platform unifying desktop, cloud, and enterprise-collaboration agents, with web development, multimodal generation, reusable skills, and DingTalk-scale ambitions behind it.
-
Tools 中阿里巴巴 QwenWork 國際版上線：全方位 workplace AI Agent 平台開放全球公測
阿里巴巴將 QwenWork 國際版開放全球公測——這是一個整合桌面、雲端與企業協作三大環境的全方位 workplace AI Agent 平台，內建網頁開發、多模態生成與可重用技能，背後還有釘釘超過 2,000 萬企業的通路優勢。
-
Models ENTencent Open-Sources Hy4 Preview: A 770B-Parameter MoE Frontier Model with 1M-Token Context
Tencent's Hunyuan team releases Hy4 preview under Apache 2.0: 770B total parameters, 49B active per token, a 1M-token context window, and benchmark scores that edge out GLM-5.3 and Kimi K3 on real-world engineering tasks.
-
Models 中騰訊開源 Hy4 preview：770B 參數 MoE 前沿模型、百萬 token 上下文
騰訊混元團隊以 Apache 2.0 授權釋出 Hy4 preview：總參數 770B、每 token 啟用 49B、上下文超過 100 萬 token，在盲測工程任務上小勝 GLM-5.3 與 Kimi K3，直攻開源前沿。
-
Industry ENDeepSeek Nears $7.4B Round at $74B Valuation as 2027 STAR Market IPO Takes Shape
DeepSeek is closing a fresh ~50 billion yuan ($7.4B) round at a ~500 billion yuan (~$74B) pre-money valuation, targeting an end-of-August close that paves the way for a possible 2026 IPO filing and a 2027 Shanghai STAR Market debut.
-
Industry 中DeepSeek 逼近以 740 億美元估值完成 74 億美元融資，2027 科創板 IPO 藍圖成形
DeepSeek 正接近完成約 500 億人民幣（74 億美元）的新一輪融資， pre-money 估值約 5,000 億人民幣（約 740 億美元），目標 8 月底完成交割，為可能於 2026 年底遞件、2027 年在上海科創板掛牌鋪路。
-
Models ENPhoneLLM Alpha 1: Open 30B Model Matches GPT-5.6 Terra on Phone Calls at 1/18th the Cost
Pipecat's PhoneLLM Alpha 1, a full fine-tune of NVIDIA's Nemotron 3 Nano 30B-A3B, matches GPT-5.6 Terra on its PhoneBench voice-agent benchmark while running 94% cheaper per minute — and it ships with no commercial restrictions.
-
Models 中PhoneLLM Alpha 1：開源 30B 模型在電話語音代理基準上追平 GPT-5.6 Terra，成本僅 1/18
Pipecat 推出的 PhoneLLM Alpha 1 是 NVIDIA Nemotron 3 Nano 30B-A3B 的全參數微調版本，在其 PhoneBench 語音代理基準上以 72.3% 追平 GPT-5.6 Terra，每分鐘成本低 94%，且不受任何商業限制。
-
Research ENSkild's S1 Learns Robot Tasks From a Single Video — No Fine-Tuning Required
Skild AI's S1 executes unseen 10-minute manipulation tasks from one human video demo, scoring 66% success versus 9% for language-prompted VLAs — the 'BERT-to-GPT-3 moment' for robotics.
-
Research ENSkild S1 只看一支影片就學會機器人任務——無需微調
Skild AI 的 S1 從單支人類示範影片就能執行長達 10 分鐘的全新操作任務，成功率 66% 對語言提示 VLA 的 9%——機器人版的「BERT 到 GPT-3 時刻」。
-
Industry ENOpenAI Is Buying Tens of Thousands of Macs for RL — and Apple Just Became an AI Hardware Company
The Information reports OpenAI has purchased tens of thousands of Macs to run reinforcement-learning workloads while Anthropic rents Mac capacity on AWS — and Nvidia now sees Apple as its principal rival in local AI processing.
-
Industry 中OpenAI 大手筆購入數萬台 Mac 跑強化學習 — 蘋果意外成為 AI 硬體公司
The Information 報導 OpenAI 已購入數萬台 Mac 執行強化學習工作負載，Anthropic 則透過 AWS 租用 Mac 運算資源；Nvidia 現在將蘋果視為本地 AI 運算的主要競爭對手。
-
Industry ENNvidia Is Buying Hugging Face for $12.9 Billion — and the Open-Source World Is Bracing
Nvidia has reportedly agreed to acquire Hugging Face, the repository hosting millions of open models, for $12.9 billion — an ~86x revenue multiple. The deal hands the chip giant the de facto home of open-source AI, and developers are asking what happens to its neutrality.
-
Industry 中Nvidia 以 129 億美元收購 Hugging Face——開源 AI 世界嚴陣以待
據報導 Nvidia 已同意以 129 億美元收購托管數百萬個開源模型的 Hugging Face，約為其營收的 86 倍。這筆交易把開源 AI 的事實上根據地交到晶片巨頭手中，開發者社群正關注平台中立性將何去何從。
-
Models ENAugust 31 Is AI's Deadline Day: Three Model Sunsets Hit at Once
Tomorrow, Claude Sonnet 5's $2/$10 intro pricing expires, GPT-5.4 leaves Codex for ChatGPT sign-in users, and Moonshot fully sunsets kimi-k2.5 and moonshot-v1 — a rare triple deadline that will reprice and reroute a huge slice of production AI traffic in a single day.
-
Models 中8 月 31 日是 AI 的截止日：三個模型日落同一天降臨
明天，Claude Sonnet 5 的 2 美元／10 美元促銷價到期，GPT-5.4 從 Codex 的 ChatGPT 登入環境退場，Moonshot 則全面下架 kimi-k2.5 與 moonshot-v1——三個獨立期限罕见地落在同一天，將在 24 小時內重新定價並改道大量生產環境的 AI 流量。
-
Industry ENTechBBQ 2026: Europe's AI Debate Shifts From 'What Can It Do' to 'Who Controls It'
At Copenhagen's TechBBQ, Europe's founders and investors argued over AI sovereignty after Anthropic's 19-day Fable and Mythos shutdown exposed the risks of renting frontier intelligence — with Signal's Meredith Whittaker warning that agentic AI is a 'data collection apparatus.'
-
Industry 中TechBBQ 2026：歐洲的 AI 討論從「能做什麼」轉向「誰來控制」
哥本哈根 TechBBQ 會議上，歐洲創業者與投資人激烈討論 AI 主權議題——Anthropic 的 Fable 與 Mythos 模型停權 19 天事件，暴露了租用前沿智慧的風險；Signal 總裁 Meredith Whittaker 更警告代理式 AI 是一部「資料蒐集機器」。
-
Tools ENvLLM 0.28.0 Lands Decode Context Parallel, DFlash2 Speculative Decoding and Tiered Disk KV Offload
The vLLM project's latest release packs 584 commits from 270 contributors: Decode Context Parallel for Kimi-K3, end-to-end sparse MLA for DeepSeek V4, DFlash2 speculation, disk-tier KV offloading, and doubled batch defaults.
-
Tools 中vLLM 0.28.0 釋出：Decode Context Parallel、DFlash2 投機解碼與磁碟層 KV 卸載全面到位
vLLM 最新版本集結 270 位貢獻者的 584 項提交：Kimi-K3 的 Decode Context Parallel、DeepSeek V4 端到端稀疏 MLA、DFlash2 投機解碼、磁碟層 KV 快取卸載，以及翻倍的批次預設值。
-
Tools ENAMD Ships ROCm 10: Agentic AI Tooling and a 3.3x Inference Claim in Its Biggest Software Release Yet
AMD's ROCm 10 marks a decade of its open compute stack with ROCm.AI — an agentic developer layer pairing ROCm CLI, AMD Skills for Claude/Cursor/Codex, and the Hyperloom auto-optimizer — plus a claimed 3.3x inference and 2.4x training uplift over ROCm 7.
-
Tools 中AMD 發布 ROCm 10：代理式 AI 工具鏈與 3.3 倍推論提升，十年來最大軟體更新
AMD 以 ROCm 10 歡慶開源運算堆疊十週年，主打 ROCm.AI 代理式開發體驗——整合 ROCm CLI、支援 Claude/Cursor/Codex 的 AMD Skills 與自動最佳化工具 Hyperloom——並宣稱推論效能較 ROCm 7 提升 3.3 倍、訓練提升 2.4 倍。
-
Models ENTencent Open-Sources Hy4 Preview: A 770B MoE Flagship With 1M-Token Context
Tencent's Hy Team has released Hy4 preview, a 770B-parameter open-weight MoE model with 49B active parameters, a 1M-token context window, and benchmark wins over GLM-5.3 and Kimi K3.
-
Models 中騰訊開源 Hy4 preview：770B 參數 MoE 旗艦模型，支援百萬 token 上下文
騰訊 Hy 團隊開源 Hy4 preview：770B 總參數、每 token 僅啟動 49B 的 MoE 旗艦模型，具備百萬 token 上下文視窗，評測成績超越 GLM-5.3 與 Kimi K3。
-
Models ENThomson Reuters Built Its Own Legal LLM for $40 Million — and It Beats GPT 5.4 on Legal Work
Thomson Reuters officially launched Thomson, a proprietary legal LLM trained on Westlaw and Practical Law content, with a final training run costing just $450K — undercutting frontier labs by orders of magnitude.
-
Models 中湯森路透只花 4,000 萬美元自建法律 LLM——在法律任務上擊敗 GPT 5.4
湯森路透正式發表自有法律大模型 Thomson，以 Westlaw 與 Practical Law 數十年的專屬內容訓練，最終訓練成本僅 45 萬美元，遠低於前沿實驗室的數十億美元投入。
-
Tools ENAnthropic's Model Hardware Standard Gives AI Agents Hands: MHS Is MCP for the Physical World
Anthropic's new Model Hardware Standard (MHS) lets AI agents safely discover, operate, and orchestrate lab and factory equipment — from Genentech liquid handlers to QuEra quantum lasers — cutting integration time from months to hours.
-
Tools 中Anthropic 推出 Model Hardware Standard：讓 AI 代理長出雙手的「實體世界版 MCP」
Anthropic 開放 Model Hardware Standard（MHS）研究預覽，讓 AI 代理能安全地探索、操作與協調實驗室及工廠設備——從 Genentech 的液體處理工作站到 QuEra 的量子雷射——整合時間從數月縮短到數小時。
-
Policy ENTrump's 'FINRA for AI' Plan Stalls: Inside the Fight Over a Frontier-Model Self-Regulator
A draft executive order creating a FINRA-style self-regulatory body for frontier AI labs has stalled inside the Trump administration, blocked partly by former AI czar David Sacks, who calls pre-release testing regimes a 'Trojan horse' against open models.
-
Policy 中川普「AI 版 FINRA」計畫卡關：前沿模型自律監管機構的幕後角力
一份建立 FINRA 式前沿 AI 自律監管機構的行政命令草案在川普政府內部停滯不前，部分阻力來自前 AI 沙皇 David Sacks——他稱發布前測試制度是對開源模型的「特洛伊木馬」。
-
Industry ENNotion Plans a 30% Hiring Binge as Ivan Zhao Goes All-In on AI Agents
The Information reports Notion will grow headcount roughly 30% this year, with new roles largely devoted to building and selling AI — the clearest signal yet that the $11B workspace app is becoming an AI company.
-
Industry 中Notion 計畫大舉擴編 30%：趙伊凡把公司全部押在 AI Agent 上
The Information 報導 Notion 今年將增加約 30% 人力，新職缺多半投入 AI 產品的開發與銷售——這是這家估值 110 億美元的協作工具公司轉型為 AI 公司最明確的訊號。
-
Models ENThe $10 Billion Clause: Z.ai's GLM-5.3 License Puts a Price Tag on Trust
GLM-5.3's weights are finally public — under a bespoke license that forces any $10B+ Model-as-a-Service provider through Z.ai's security review. As US labs reel from a summer of AI security scares, China is pitching openness itself as the safer bet.
-
Models 中100 億美元條款：Z.ai 的 GLM-5.3 授權為「信任」標上價格
GLM-5.3 權重終於公開，但採用了一份量身打造的授權：年營收超過 100 億美元的模型服務業者，必須先通過 Z.ai 的安全審查。在美國實驗室飽受一連串 AI 安全事件衝擊之際，中國正把「開放」本身包裝成更安全的選擇。
-
Models ENQwen3.8-Flash-Next: Alibaba Open-Sources the First Glimpse of Qwen4's Architecture
Alibaba's Qwen team open-weights a 125B-parameter MoE that activates just 6B per token, pairs 1M-token context with a hybrid GDN + QSA attention stack, and beats Claude Opus 4.6 Max on agentic coding benchmarks.
-
Models 中Qwen3.8-Flash-Next：阿里巴巴開源 Qwen4 架構的首波預覽
阿里巴巴 Qwen 團隊開源一款 125B 參數的 MoE 模型，每 token 僅啟動 6B，兼顧百萬級上下文與 GDN + QSA 混合注意力，在代理式編程基準上超越 Claude Opus 4.6 Max。
-
Research ENAn Extra Day of Warning: Google DeepMind Open-Sources WeatherNext, the AI That Jumped Hurricane Forecasting a Decade Ahead
Google DeepMind's WeatherNext models predict a cyclone's track, intensity and wind structure a full day earlier than conventional systems — a decade of meteorological progress in one model, now open-sourced.
-
Research 中多贏一天預警:Google DeepMind 開源 WeatherNext,讓颶風預報一口氣躍進十年
Google DeepMind 的 WeatherNext 模型預測氣旋路徑、強度與風場結構,比傳統系統平均提早整整一天——相當於把氣象預測進度一次推進十年,而且程式碼與權重已全面開源。
-
Industry ENNvidia Compresses AI Model Release Cycle to 4–6 Weeks — Software Sprints While Hardware Walks
Nvidia's VP of Applied Deep Learning Bryan Catanzaro says the company now ships open-weight model updates every 4–6 weeks, down from 6–8 months — a 5–8x acceleration that applies to software only, while Blackwell and Vera Rubin keep their annual silicon cadence.
-
Industry 中Nvidia 將 AI 模型發布週期壓縮至 4–6 週 — 軟體衝刺，硬體慢走
Nvidia 應用深度學習研究副總裁 Bryan Catanzaro 證實，開放權重模型的更新週期已從每 6–8 個月縮短至每 4–6 週——約 5–8 倍的加速，但僅限軟體；Blackwell 與 Vera Rubin 矽晶片仍維持年度節奏。
-
Models ENFive Open-Weight Models in Nine Days: The Capability Premium Just Collapsed
Between August 21 and 29, five Chinese AI labs shipped open-weight models with 1M-token context at budget prices — and OpenAI answered with a 20% price cut. The frontier is now a price war.
-
Models 中九天五款開源權重模型：能力溢價正在崩塌
8 月 21 日至 29 日，五家中國 AI 實驗室接連推出具備百萬 token 上下文的開源權重模型，價格卻壓在低價帶——OpenAI 的回應是降價 20%。前沿模型市場正式進入價格戰。
-
Industry ENOpenAI Cuts Cursor Off: GPT Models Vanish November 12 After SpaceX Takeover
OpenAI is winding down the contract that supplies GPT models to Cursor, with a proposed shutoff of November 12, 2026 — the first big fallout of SpaceX's $60 billion acquisition of the coding agent.
-
Industry 中OpenAI 切斷 Cursor 模型供應：SpaceX 收購後首個重大商業衝擊
OpenAI 宣布終止向 Cursor 供應 GPT 模型的合約，提議關閉日期為 2026 年 11 月 12 日——這是 SpaceX 以 600 億美元收購這款 AI 編程代理後引發的第一場重大餘波。
-
Industry ENMeta Readies 'Hatch' AI Agent and October's 'Watermelon' Model in a Consumer Monetization Push
Meta will launch its consumer AI agent platform Hatch within weeks — with a premium tier reportedly priced up to $199.99 a month — followed by the Watermelon frontier model in October, the clearest signal yet that free Meta AI is giving way to paid subscriptions.
-
Industry 中Meta 備戰消費級 AI 變現：代理平台「Hatch」數週內登場，十月推出「Watermelon」前沿模型
Meta 將在數週內推出消費級 AI 代理平台 Hatch——據報最高階方案月費達 199.99 美元——十月再發布前沿模型 Watermelon，這是「免費 Meta AI」讓位給付費訂閱迄今最明確的訊號。
-
Models ENTencent Open-Sources Hy4 Preview: A 770B MoE Flagship Built for Real Work
Tencent's Hunyuan team releases Hy4 preview, a 770B-parameter MoE model with 49B active parameters, 1M-token context and an early recursive self-improvement loop.
-
Models 中騰訊開源 Hy4 Preview：770B 參數 MoE 旗艦模型，還學會了自我改進
騰訊混元團隊發布 Hy4 preview：770B 總參數、49B 激活參數的 MoE 旗艦模型，具備百萬 token 上下文與早期遞迴自我改進迴路，採 Apache 2.0 開源。
-
Industry ENNvidia Agrees to Buy Hugging Face for $12.9 Billion — and the Open-Source World Is Bracing for Impact
Nvidia has agreed to acquire Hugging Face, the 'GitHub of AI' that hosts millions of open models and the llama.cpp inference engine, for $12.9 billion — putting a single chipmaker in control of open-source AI's most important commons.
-
Industry 中輝達同意以 129 億美元收購 Hugging Face——開源 AI 世界正嚴陣以待
輝達已同意以 129 億美元收購被譽為「AI 界 GitHub」的 Hugging Face，連同其託管的數百萬個開源模型與 llama.cpp 推理引擎——這意味著單一晶片廠商將掌控開源 AI 最重要的公共財。
-
Research ENClaude Aligns Claude: Anthropic's Automated Researchers Beat Human Safety Experts at Fixing Misaligned AI
Anthropic's new paper shows Claude autonomously running alignment research — closing 85% of the deception safety gap where human experts closed 20%, and post-training an early Opus 4.8 checkpoint with a recipe 15,000x more efficient than production.
-
Research 中Claude 為 Claude 對齊：Anthropic 自動化研究員修復 AI 失準問題，表現超越人類安全專家
Anthropic 最新論文顯示，Claude 能自主執行對齊研究——在欺騙行為上關閉 85% 的安全差距（人類專家僅 20%），並在 60 小時內以比生產流程快 15,000 倍的配方，為早期 Opus 4.8 檢查點完成對齊後訓練。
-
Models ENDeepSeek's V4-Flash-Vision-Exp Sees for Pennies: The 384-Token Gamble Reshaping Agent Economics
DeepSeek's experimental multimodal model matches V4-Flash text performance at identical prices, caps every image at 384 tokens, and trails Opus-4.8 by single points on multimodal agent benchmarks — while coding harnesses burn billions of tokens through it in days.
-
Models 中DeepSeek V4-Flash-Vision-Exp：每張圖 384 Token 的視覺賭注，正在改寫 Agent 經濟學
DeepSeek 的實驗性多模態模型以與 V4-Flash 完全相同的價格提供視覺能力，每張圖片固定計費 384 token，多模態 Agent 基準逼近 Opus-4.8——上線三天就被編碼代理燒掉 250 億 token。
-
Models ENTencent Open-Sources Hy4 Preview: 770B MoE Flagship With 1M Context and a 31.8% Self-Tuning Speedup
Tencent Hunyuan's new open-weight flagship packs 770B total / 49B active parameters, a 1M-token window, and an early recursive self-improvement loop that lifted its own inference throughput 31.8%.
-
Models 中騰訊開源 Hy4 Preview：770B MoE 旗艦、百萬 token 上下文，還親手把自己的推理吞吐提速 31.8%
騰訊混元新一代開源旗艦 Hy4 preview 擁有 770B 總參數、49B 激活參數與超過 1M token 的上下文視窗，並首度建立遞迴自我改進迴路，自行優化推論系統帶來 31.8% 吞吐提升。
-
Research ENMIT's η-Learning Generates Extreme-Event Scenarios Without Ever Seeing One
MIT's Extreme Event Aware (η-learning) algorithm generates plausible maps of unprecedented storms, floods, and wildfires without training on historical disasters — published Aug 20 in Nature Communications.
-
Research 中MIT 的 η-Learning：從未見過極端事件，也能生成極端事件場景
MIT 團隊的 Extreme Event Aware（η-learning）演算法，無需以歷史災害資料訓練，就能生成前所未見的暴雨、洪水與野火場景地圖——論文於 8 月 20 日刊於 Nature Communications。
-
Models ENZ.ai's GLM-5.3-Flash: Frontier Multimodal Intelligence at One-Tenth the Cost
Z.ai open-sources GLM-5.3-Flash, a 320B-parameter hybrid-attention multimodal model that matches Claude Opus 4.8 on coding at roughly a tenth of the price — with 1M-token context, MIT license, and inference served at scale on Chinese AI chips.
-
Models 中Z.ai 的 GLM-5.3-Flash：十分之一成本的前沿多模態模型
Z.ai 開源 GLM-5.3-Flash：320B 參數混合注意力多模態模型，程式能力接近 Claude Opus 4.8、價格僅約十分之一，支援百萬 token 上下文、MIT 授權，並大規模部署於中國自研 AI 晶片上。
-
Tools ENVisa's Security AI Now Patches Production Code Before Any Human Reviews It
Visa's open-source VVAH harness now discovers, patches, and adversarially validates vulnerabilities in one autonomous loop — with human review pushed to the edges of the pipeline.
-
Tools 中Visa 的安全 AI 開始在人類審查之前直接修補生產環境程式碼
Visa 開源的 VVAH 安全框架升級後，能在單一自主循環中完成漏洞發現、修補與對抗性驗證——人類審查被推向管線的兩端。
-
Models ENGLM-5.3 Open Weights Land: Z.ai's Exploit-Hunting Flagship Finally Goes Public
After a two-week safety review, Z.ai's GLM-5.3 open weights are scheduled to hit Hugging Face today — and GLM-5.3-Flash's MIT-licensed weights are already live, turning every self-reported benchmark claim into something anyone can verify.
-
Models 中GLM-5.3 開放權重上線：Z.ai 那個會找漏洞的旗艦模型終於公開
歷經兩週安全審查，Z.ai 的 GLM-5.3 開放權重預計今（28）日登上 Hugging Face，而 MIT 授權的 GLM-5.3-Flash 權重已經上線——所有自報的基準分數，從今天起人人都可以親自驗證。
-
Industry ENChina Now Runs Over 2 Million Factory Robots — and Half the World Makes Them
A BBC on-the-ground report finds more than 2 million industrial robots at work in Chinese factories, fueled by a shrinking workforce and a $20 billion state plan.
-
Industry 中中國工廠機器人突破 200 萬台——全球一半機器人都是它造的
BBC 實地報導揭露：中國工廠在役工業機器人已超過 200 萬台居全球之冠，背後是人口萎縮與 200 億美元的國家自動化戰略。
-
Research ENSkild AI's S1 Learns New Robot Tasks From a Single Video — No Fine-Tuning Required
Skild AI's S1 robotics foundation model executes 10-minute unseen manipulation tasks from one video demonstration, hitting 66% step success versus 9% for language-prompted VLAs.
-
Research 中Skild AI 的 S1：看一支影片就學會新任務的機器人基礎模型，無需微調
Skild AI 發布機器人基礎模型 S1，僅憑一支影片示範就能執行長達 10 分鐘、訓練時從未見過的操作任務，步驟成功率 66%，遠勝語言提示 VLA 的 9%。
-
Tools ENHugging Face's Microduck: The $399 Open-Source Robot Duck You Train With Reinforcement Learning
Hugging Face opens pre-orders for Microduck, a $399, 25cm bipedal robot duck with 15 motors, LiDAR, a grasping beak, and a full open-source RL training stack — ships before Christmas 2026.
-
Tools 中Hugging Face 推出 Microduck：399 美元的開源機器鴨，用強化學習教會牠新把戲
Hugging Face 開放 Microduck 預購——399 美元、25 公分高的雙足機器鴨，內建 15 顆馬達、LiDAR 與可抓取的鴨嘴，並附上完整開源的強化學習訓練工具鏈，目標 2026 年聖誕節前出貨。
-
Industry ENNvidia Buys the GitHub of AI: Inside the $12.9 Billion Hugging Face Deal
Nvidia has agreed to acquire Hugging Face — the repository hosting millions of open-weight models — for $12.9 billion, wrapping its arms around the central hub of the open-source AI ecosystem.
-
Industry 中Nvidia 買下 AI 界的 GitHub：129 億美元併購 Hugging Face 全解析
Nvidia 同意以 129 億美元收購 Hugging Face——這個託管數百萬個開放權重模型的平台，等於把開源 AI 生態系的中樞整整抱進懷裡。
-
Research ENA 27B 'AI Scientist' That Directs GPT-5.5: Inside Inherent's Faraday and the Replica Benchmark
London lab Inherent shows a 27-billion-parameter agent post-trained with long-horizon RL can out-replicate Claude Opus 4.8 and GPT-5.5 — by learning to direct the very frontier models it beats, on a 310-task benchmark where the reward is scientific taste, not code.
-
Research 中會指揮 GPT-5.5 的 270 億參數「AI 科學家」：Inherent 的 Faraday 與 Replica 基準深度解析
倫敦新創 Inherent 證明：用長程強化學習後訓練的 270 億參數代理，能在論文重現任務上擊敗 Claude Opus 4.8 與 GPT-5.5——靠的不是自己寫程式，而是學會指揮那些比自己大上幾個數量級的前沿模型。
-
Industry ENMoonshot AI Wants 30% of Hyperscaler Revenue to Host Kimi K3 — and It Might Have the Leverage
Reuters reports that China's Moonshot AI is in early talks with Microsoft, Amazon, and Google to host Kimi K3 on their clouds — demanding up to 30% of K3-related revenue, a deal that would invert the app-store model and test how much US hyperscalers need frontier open-weight models.
-
Industry 中Moonshot AI 要求三大雲端巨頭上繳 30% 營收才能託管 Kimi K3——而它可能真有談判籌碼
路透社獨家報導，中國 Moonshot AI 正與微軟、亞馬遜、Google 早期洽談，要求在 Azure、AWS、Google Cloud 上託管 Kimi K3 的營收分成最高可達 30%——這樁交易若成，將是首件中國 AI 實驗室與美國雲端巨頭的營收共享協議，並徹底顛覆 app store 式的平台抽成邏輯。
-
Models ENOx Alpha Unmasked: Z.ai's Stealth Model Is GLM-5.3-Flash, a 320B Open-Weight Multimodal Contender
The anonymous model that topped OpenRouter's usage charts turned out to be Z.ai's GLM-5.3-Flash — a 320B-A18B natively multimodal MoE with 1M context and MIT-licensed weights at a tenth of frontier pricing.
-
Models 中Ox Alpha 揭曉真身：Z.ai 的匿名模型就是 GLM-5.3-Flash——320B 開源多模態強權
登上 OpenRouter 用量冠軍的匿名模型「Ox Alpha」，證實是 Z.ai（智譜）的 GLM-5.3-Flash——320B-A18B 原生多模態 MoE、百萬 token 上下文、MIT 授權開放權重，價格僅前線模型的十分之一。
-
Policy ENX Shuts Down Nitter With Cease-and-Desist — and the Open-Source Code May Be Next
X Corp's legal takedown of the seven-year-old Nitter project removes a key open window into public posts — one that AI agents, researchers and privacy tools had quietly come to depend on.
-
Policy 中X 以存證信函終結 Nitter——開源程式碼本身恐是下一個目標
X Corp 的法律行動終結了這個七年的開源專案，也關上了一扇觀看公開貼文的窗——而 AI 代理、研究人員與隱私工具，早已默默依賴這扇窗。
-
Industry ENNvidia Nears $12.9B Hugging Face Acquisition in Bid to Own Open AI's Home Turf
Nvidia has reportedly agreed to buy Hugging Face for $12.9 billion — a 3x jump from its 2023 valuation — to anchor its open-source AI ecosystem play as closed labs design their own chips.
-
Industry 中Nvidia 以 129 億美元收購 Hugging Face：晶片霸主買下開源 AI 的根據地
據報導 Nvidia 已同意以 129 億美元收購 Hugging Face，估值較 2023 年暴增近三倍，藉此鞏固開源 AI 生態系，對抗自研晶片的封閉實驗室。
-
Models ENQwen3.8-Flash-Next: Alibaba Opens the Door on the Qwen4 Architecture
Alibaba's Qwen team releases a 125B-parameter open-weight preview of the Qwen4 architecture — hybrid linear attention, a 51B-parameter n-gram embedding, and 6B active parameters per token.
-
Models 中Qwen3.8-Flash-Next：阿里巴巴提前揭開 Qwen4 架構的面紗
阿里巴巴 Qwen 團隊發布 125B 參數的開放權重模型，預覽 Qwen4 架構：混合線性注意力、51B 參數的 n-gram 嵌入層，以及每 token 僅 6B 活躍參數。
-
Models ENMystery Solved: Ox Alpha Was GLM-5.3-Flash All Along — 320B Open Weights for $0.15/M Tokens
Z.ai unmasked its anonymous stealth model: GLM-5.3-Flash is a 320B-A18B natively multimodal MoE with 1M context and MIT-licensed weights at $0.15/$0.50 per million tokens — and the entire preview ran on Chinese-made AI chips.
-
Models 中謎底揭曉：Ox Alpha 就是 GLM-5.3-Flash——320B 開放權重、每百萬 token 只要 0.15 美元
Z.ai 揭曉匿名 stealth 模型的真實身分：GLM-5.3-Flash 是 320B-A18B 原生多模態 MoE，具備 1M context 與 MIT 授權權重，API 定價每百萬 token 輸入 0.15／輸出 0.50 美元——而且整個預覽期間都跑在中國自產 AI 晶片上。
-
Industry ENAWS Acquires DuckLabs: The Team Behind DuckDB Joins Amazon, Open Source Project Stays Independent
Amazon has signed a definitive agreement to acquire DuckLabs, the Amsterdam company behind the wildly popular open source analytical database DuckDB. The DuckDB project itself stays MIT-licensed under the independent DuckDB Foundation — but the deal hands AWS the engineering brains behind one of the fastest-adopted data tools of the AI era.
-
Industry 中AWS 收購 DuckLabs：DuckDB 背後團隊加入 Amazon，開源專案維持獨立
Amazon 已簽署最終協議，收購開源分析資料庫 DuckDB 背後的阿姆斯特丹公司 DuckLabs。DuckDB 專案本身仍由獨立的 DuckDB Foundation 以 MIT 授權管理——但這筆交易讓 AWS 把 AI 時代成長最快的資料工具之一的工程大腦納入麾下。
-
Models ENIBM's Granite 4.2 Brings Open-Weight Reasoning to Enterprise Agents
IBM's new 3B/8B/30B open-weight models add native thinking modes and agentic RL training, aiming reasoning at on-prem enterprise workloads.
-
Models 中IBM Granite 4.2 將開放權重推理能力帶進企業 Agent
IBM 發布 3B/8B/30B 開放權重模型，加入原生思考模式與 Agentic RL 訓練，把推理能力推向在地部署的企業工作負載。
-
Models ENOx Alpha Revealed: Z.ai's GLM-5.3-Flash Ships Open Weights, Frontier Scores, and a Chinese-Chip Serving Stack
Z.ai confirmed the stealth model Ox Alpha is GLM-5.3-Flash: a 320B-parameter MIT-licensed multimodal model with near-frontier coding scores, a 1M-token context window, aggressive pricing, and inference served on tens of thousands of domestic Chinese AI chips.
-
Models 中Ox Alpha 揭曉：Z.ai 的 GLM-5.3-Flash 開放權重、前線級分數與中國晶片推理叢集
Z.ai 證實 stealth 神秘模型 Ox Alpha 就是 GLM-5.3-Flash：320B 參數、MIT 授權開放權重的原生多模態模型，具備百萬 token 上下文與逼近前線的編碼分數，且整個曝光週期的推理服務都跑在數萬顆中國國產 AI 晶片上。
-
Tools ENEx-Meta FAIR Scientists Open-Source Isaac 0.5, a Vision Model That Lets Industrial Robots Perceive, Reason, and Act
Perceptron, a startup founded by two former Meta FAIR researchers, has released Isaac 0.5 — an open-weight vision model trained on a million hours of video, built to guide robots through warehouses and factory floors without cloud GPUs.
-
Tools 中前 Meta FAIR 科學家開源 Isaac 0.5：讓工業機器人能感知、推理與行動的視覺模型
由兩位前 Meta FAIR 研究科學家創辦的 Perceptron 發布了 Isaac 0.5——以百萬小時影片訓練的開放權重視覺模型，無需雲端 GPU 即可引導機器人穿越倉庫與工廠。
-
Industry ENMiniMax Nearly Quadruples Revenue in First Post-IPO Interim Results, as Enterprise API Demand Explodes 703%
MiniMax reported H1 2026 revenue of US$116.6 million, up 283% year over year and already exceeding all of 2025, with Open Platform enterprise revenue surging 703% as token consumption hit 20x its January level.
-
Industry 中MiniMax 上市後首份中期財報營收近翻四倍，企業 API 業務暴增 703%
MiniMax 公布 2026 上半年營收 1.166 億美元、年增 283%，已超越 2025 全年總和；Open Platform 企業業務暴增 703%，Token 消耗量更達 1 月的 20 倍。
-
Industry ENDeepSeek's Revenue Jumped Tenfold to $70.7M — and Its API Business Is More Profitable Than OpenAI's
Leaked financials show DeepSeek generated RMB 475M ($70.7M) in the first seven months of 2026, a tenfold jump, with API gross margin at a remarkable 82.9%.
-
Industry 中DeepSeek 營收暴增十倍至 7,070 萬美元——API 毛利率竟超越 OpenAI
外流財報顯示，DeepSeek 2026 年前七個月營收達 4.75 億人民幣（7,070 萬美元），年增十倍，API 毛利率更高達 82.9%。
-
Industry ENEmerald AI Raises $150M at $1.05B Valuation to Turn AI Data Centers Into Grid Assets
Emerald AI's oversubscribed Series A, backed by NVIDIA, Siemens and GE Vernova, bets that software orchestration — not new power plants — is the fastest way through AI's electricity bottleneck.
-
Industry 中Emerald AI 以 10.5 億美元估值融資 1.5 億美元,要讓 AI 資料中心變成電網的彈性資產
獲得 NVIDIA、西門子、GE Vernova 等巨頭共同投資的超額認購 A 輪融資,押注軟體協調——而非新建電廠——才是突破 AI 電力瓶頸的最快解方。
- Industry EN
Two Million Robots and Counting: Inside China's Factory Automation Sprint Against Its Own Demographics
A BBC on-the-ground report finds more than two million industrial robots at work in Chinese factories — over half the world's supply is now built in China — as Beijing races to automate before an aging, shrinking workforce catches up with its economy.
- Industry 中
兩百萬台機器人還在增加:透視中國工廠自動化與人口老化的賽跑
BBC 實地報導發現,中國工廠裡已有超過兩百萬台工業機器人運轉——全球一半以上的工業機器人由中國製造——北京正與時間賽跑,希望在人口老化與萎縮拖垮經濟之前完成自動化轉型。
-
Research ENMIT's η-Learning Generates Extreme-Weather Scenarios That Never Happened — Yet
MIT's Extreme Event Aware (η-learning) algorithm invents statistically plausible once-in-a-century storms without ever training on historical disasters.
-
Research 中MIT 的 η-learning 演算法：憑空生成史上從未發生過的極端天氣情境
MIT 開發的 Extreme Event Aware（η-learning）演算法，不需要任何歷史災害資料，就能產出統計上合理的百年一遇風暴地圖。
-
Models ENMeta's Comeback Play: Muse Code, Muse Spark 1.2, and the Open-Weights Muse Glimmer
Meta Superintelligence Labs has shipped Muse Code (beta), a terminal coding agent powered by Muse Spark 1.2, and open-sourced the 30B Muse Glimmer under Apache 2.0 — the company's full return to open weights after its closed-API pivot.
-
Models 中Meta 的絕地反攻：Muse Code、Muse Spark 1.2 與開放權重的 Muse Glimmer
Meta 超級智慧實驗室發布了終端編碼代理 Muse Code（beta）、背後的 Muse Spark 1.2 模型，並以 Apache 2.0 授權開源 30B 的 Muse Glimmer —— 這是 Meta 在轉向閉源 API 之後，正式重返開放權重陣營。
-
Models ENNobody Knows Who Built Ox Alpha: The Anonymous Model Beating GPT-5.6 at Coding
A frontier-class stealth model called Ox Alpha appeared free on OpenRouter on August 20 with a 1M-token context window, video input, and an 80% DeepSWE score that tops Claude Fable 5 and GPT-5.6 Sol — and its creator is staying anonymous.
-
Models 中沒人知道誰做了 Ox Alpha：擊敗 GPT-5.6 的匿名模型
一款名為 Ox Alpha 的前沿級隱身模型於 8 月 20 日免費現身 OpenRouter，具備百萬 token 上下文、視訊輸入，以及以 80% 的 DeepSWE 分數壓過 Claude Fable 5 與 GPT-5.6 Sol——而它的創造者選擇匿名。
-
Meta ENOne Malicious Webpage Can Now Silently Poison Your Local AI Agent: Inside NVIDIA NemoClaw's CVE-2026-65105
Oasis Security details CVE-2026-65105: NemoClaw binds Ollama to 0.0.0.0:11434 with no auth, so a DNS-rebinding webpage can rewrite the model's chat template and steer a developer's AI agent persistently, invisibly, from the browser.
-
Industry ENStability AI's $76M Series B: All Three Major Labels and EA Now Own Stakes in the Stable Diffusion Maker
Universal, Sony, and Warner Music plus Electronic Arts, AMD Ventures and Pacific Alliance Ventures have joined Stability AI's $76M Series B, bringing the company's total funding under CEO Prem Akkaraju to $232M — the clearest sign yet that rights holders now want equity in the AI tools built on their catalogs.
-
Industry 中Stability AI 完成 7,600 萬美元 B 輪融資：三大唱片與 EA 齊入股 Stable Diffusion 母公司
環球、索尼、華納三大音樂集團攜手 Electronic Arts、AMD Ventures 與 Pacific Alliance Ventures 参与 Stability AI 的 7,600 萬美元 B 輪融資，使该公司在執行長 Prem Akkaraju 任內總募資達 2.32 億美元——版權方從提告者變成股東，成為娛樂產業擁抱生成式 AI 最明確的訊號。
-
Models ENAlibaba Previews the Qwen4 Architecture Early: Qwen3.8-Flash-Next Open Weights Drop August 26
Hours before the weights go live, Alibaba is shipping its Qwen4 architecture — GDN hybrid attention plus Qwen Sparse Attention — inside an open-weight multimodal MoE branded Qwen3.8-Flash-Next, rumored at 125B total / 6B active parameters.
-
Models 中阿里巴巴提前公開 Qwen4 架構：Qwen3.8-Flash-Next 開放權重 8 月 26 日登場
在權重上架前的幾小時，阿里巴巴將下一代 Qwen4 架構——GDN 混合注意力加上 Qwen Sparse Attention——包進開放權重的多模態 MoE 模型 Qwen3.8-Flash-Next 釋出，傳聞總參數 125B、激活參數 6B。
-
Models ENSkild AI's S1 Learns 10-Minute Robot Tasks From a Single Video — No Fine-Tuning
Skild AI's S1 robotics foundation model executes never-before-seen manipulation tasks up to 10 minutes long from one video demonstration — 66% success on unseen tasks with zero fine-tuning, 7x better than language prompting.
-
Models 中Skild AI 發表 S1：看一支影片就學會 10 分鐘長任務，完全不需微調
Skild AI 的機器人基礎模型 S1 只需一支影片示範，就能執行訓練時從未見過、長達 10 分鐘的操作任務——未見任務成功率 66%，零微調，比語言提示高出 7 倍。
-
Policy ENAnthropic Puts $5 Million Behind Independent Research on AI's Impact on User Wellbeing
Anthropic launched a $5 million grant program on August 25, 2026 to fund fully independent, open-source evaluations of how AI models affect user wellbeing — applications close September 21, with full-proposal invites going out October 5.
-
Policy 中Anthropic 投入 500 萬美元，資助獨立研究 AI 對使用者福祉的影響
Anthropic 於 2026 年 8 月 25 日宣布啟動 500 萬美元補助計畫，資助完全獨立、開源的研究，評測 AI 模型對使用者福祉的影響——申請截止於 9 月 21 日，10 月 5 日通知入選者提交完整提案。
-
Tools ENMeta's Hatch Goes Live in Weeks: Watermelon Model Slated for October as the Agent Platform Takes Shape
The Information reports Meta will launch its Hatch consumer AI agent platform in late August or early September, with the GPT-5.5-class 'Watermelon' model following in October and a WhatsApp third-party agent trial starting within days.
-
Tools 中Meta 的 Hatch 數週內上線：Agent 平台成形、Watermelon 模型十月登場
據 The Information 報導，Meta 將於八月底或九月初推出消費級 AI agent 平台 Hatch，十月再推出代號 Watermelon 的新前沿模型，WhatsApp 的第三方 agent 試驗最快本週展開。
-
Research ENMIT's η-Learning Generates Worst-Case Weather Scenarios Without Ever Seeing One
MIT's Extreme Event Aware (η-) learning generates plausible once-in-a-century storms, floods, and wildfires without training on historical disasters — published in Nature Communications.
-
Research 中MIT 的 η-學習：不必見過災難，也能生成百年一遇的極端天氣情境
MIT 團隊在 Nature Communications 發表「極端事件感知學習」（η-learning），無需歷史災害資料即可生成百年一遇暴雨、洪水與野火的合理最壞情境。
-
Tools ENAnthropic Takes Its Agent Stack to General Availability: Computer Use, Browser Tool, Skills and Files APIs Go Production
On August 20, 2026, Anthropic moved computer use, the new browser tool, the Skills API, and the Files API to general availability on the Claude Platform — the production stack for building agents, with no beta header required.
-
Tools 中Anthropic 代理工具全面轉正：Computer Use、瀏覽器工具、Skills 與 Files API 正式上線
2026 年 8 月 20 日,Anthropic 將 computer use、全新的瀏覽器工具、Skills API 與 Files API 一次性推向正式版(GA)——這是 Claude 平台上打造 AI 代理的完整生產環境,不再需要任何 beta 標頭。
-
Models ENApple's M6 and M5 Ultra: First 2nm Chip and Quad-Die Monster Aim squarely at On-Device AI
Apple's M6 is its first 2nm chip with a Dual 16-core Neural Engine, while M5 Ultra fuses four dies into the most powerful M-series SoC ever — 512GB unified memory, 1.2TB/s bandwidth, built to run frontier AI models locally.
-
Models 中Apple M6 與 M5 Ultra：首款 2nm 晶片與四晶粒巨獸，正面搶攻裝置端 AI
Apple M6 是首款 2nm 製程晶片、配備雙 16 核心神經網路引擎；M5 Ultra 則以四晶粒架構成為史上最強 M 系列晶片——512GB 統一記憶體、1.2TB/s 頻寬，目標是讓前沿 AI 模型直接在本機跑。
-
Industry ENHugging Face's Revenue Jumped 50% to $150M — and a $13B Sale Is Now 'Nearing'
The Information reports Hugging Face's annualized revenue surged 50% to $150 million in two months even as the open-model hub nears a deal to sell itself at $13B+. Growth and exit, at the same time.
-
Industry ENOpenAI-Backed Harvey Builds Its First In-House Model on China's Kimi K3
Legal AI leader Harvey post-trained Harvey Tenet on Moonshot AI's open-weight Kimi K3 — the clearest sign yet that Chinese open models are becoming the foundation layer for Western enterprise AI.
-
Industry 中OpenAI 投資的 Harvey，第一款自研模型選擇了中國的 Kimi K3 作為基底
法律 AI 龍頭 Harvey 以月之暗面的開源權重模型 Kimi K3 為基底，後訓練出首款自研模型 Harvey Tenet——這是中國開源模型成為西方企業 AI 基礎層最明確的訊號。
-
Models ENMiniMax Open-Sources Music 3: Complete Five-Minute Songs From One Prompt
MiniMax Music 3 pairs an 8B Global LLM with a 0.6B Local LLM and a flow-matching synth stage to generate full five-minute songs — and the weights are on Hugging Face.
-
Models 中MiniMax 開源 Music 3：一次生成完整五分鐘歌曲
MiniMax Music 3 結合 8B 全域 LLM、0.6B 區域 LLM 與流匹配合成階段，可生成完整五分鐘歌曲——模型權重已上架 Hugging Face 供所有人下載。
-
Models ENGoogle's Gemini 3.7 Flash: Near-Frontier Coding Performance at Half the Price
Three weeks after 3.6 Flash, Google ships Gemini 3.7 Flash — a workhorse model for coding and agents that jumps DeepSWE from 49% to 65.3% and FrontierCode from 34.4% to 43.6%, at half the intro price of its predecessor.
-
Models 中Google Gemini 3.7 Flash：以半價提供接近前沿的程式碼能力
距離 3.6 Flash 僅三週，Google 就推出 Gemini 3.7 Flash——專為程式開發與 Agent 而生的主力模型，DeepSWE 從 49% 躍升至 65.3%、FrontierCode 從 34.4% 升至 43.6%，入門價卻只有前代的一半。
-
Industry ENNvidia Pays Poolside $6 Billion for Its 'Model Factory' — and Takes 109 Engineers With It
Nvidia is paying $6B to license Poolside's Model Factory, investing $1B more at a $12B valuation, and hiring 109 engineers to supercharge its open-weight Nemotron project — Washington's answer to DeepSeek and Qwen.
-
Industry 中Nvidia 斥資 60 億美元取得 Poolside「模型工廠」連人帶軟體一起搬
Nvidia 以 60 億美元非獨家授權 Poolside 的 Model Factory，另投資 10 億美元、估值 120 億美元，並挖走 109 位工程師投入開源 Nemotron——這是美國陣營對 DeepSeek 與 Qwen 的正面回擊。
-
Models ENApodex 1.1 Pitches 'Environment Scaling' for Agents and Ships a 35B Mini You Can Run Locally
A 70-author paper claims two new scaling axes — executable environments and agentic coordination — push a 35B open-weights model into frontier territory on finance and science benchmarks.
-
Models 中Apodex 1.1 提出「環境規模化」代理人新路線，同步開源可在本地部署的 35B Mini 模型
一篇 70 位作者共同掛名的論文主張：代理能力的下一波躍升來自規模化「可執行環境」與「多代理人協作」，並以 35B 開源權重模型在財務與科學基準闖進前沿地帶。
-
Industry ENFrom Fortnite to Four Legs: General Intuition's Value Nearly Triples to $6B on the 'Large Action Model' Bet
Valor Equity, Point72 Ventures and Seven Seven Six are backing world-model startup General Intuition at a $6B pre-money valuation — nearly triple its $2.3B mark from just eight weeks ago — as the Medal spin-out pushes its action-labeled gameplay models into robotic embodiments.
-
Industry 中從《要塞英雄》到機器狗：General Intuition 搶攻「大型動作模型」，估值八週暴增至 60 億美元
Valor Equity、Point72 Ventures 與 Seven Seven Six 將以 60 億美元投前估值投資世界模型新創 General Intuition——是八週前 23 億美元估值的近三倍——這家從遊戲錄影平台 Medal 分拆的公司正把動作標註遊戲資料模型推向機器人實體應用。
-
Tools ENLaude Institute Open-Sources Headlong: A Persistent AI Agent That Never Stops Thinking
A sub-10K-line Bash 'microharness' keeps an LLM in a self-guided inner-monologue loop around the clock — for $1-2 an hour — and the lab's agent Audel has already shipped 50+ commits back into the codebase.
-
Tools 中Laude 研究所開源 Headlong：一個永不停止思考的持續型 AI Agent
這個不到 1 萬行 Bash 程式碼的「微型 harness」讓 LLM 全天候處於自我引導的內心獨白迴圈——每小時只要 1 到 2 美元——實驗室的 agent「Audel」已經把 50 多個 commit 貢獻回程式庫。
-
Models ENOx Alpha: The Anonymous Frontier Model Nobody Can Trace
A stealth AI model with 1M-token context appeared free on OpenRouter and started beating GPT-5.6 and Claude on coding benchmarks. Nobody knows who built it — but the fingerprints are getting clearer.
-
Models 中Ox Alpha：沒有人能追溯的匿名前沿模型
一個擁有百萬 token 上下文視窗的隱身 AI 模型免費現身 OpenRouter，並在程式碼基準測試中擊敗 GPT-5.6 與 Claude。沒有人知道它是誰打造的——但指紋線索正越來越清晰。
-
Models ENThomson Reuters Launches 'Thomson', a $40 Million Frontier Model Trained on Decades of Legal and Tax Content
The legal-tech giant built its own LLM from an open-source base for a fraction of frontier-lab cost — and early benchmarks put it alongside Claude Opus 4.8.
-
Models 中湯森路透推出自研前沿模型 Thomson：4,000 萬美元打造的法律稅務專用 LLM
這家法律科技巨頭以開源模型為基礎，用不到前沿實驗室零頭的成本自研 LLM，早期評測已與 Claude Opus 4.8 並駕齊驅。
-
Policy ENDeepSeek Is Now the Weapon of Choice for Chinese State Hackers — And It Doubled Their Attack Volume
TeamT5 says Chinese state-affiliated groups have more than doubled attack volume by wiring DeepSeek into reconnaissance, exploit writing, and lateral movement — drawn by weak cyber guardrails and rock-bottom costs.
-
Policy 中DeepSeek 成為中國駭客國家隊的新武器——攻擊量直接翻倍
台灣資安廠商 TeamT5 提出報告：中國官方背景駭客組織將 DeepSeek 導入偵察、漏洞攻擊與橫向移動流程後，攻擊量已翻倍——低廉成本與寬鬆防護欄是主因。
-
Industry ENNvidia Pays Poolside $6 Billion for Its 'Model Factory' — the Largest AI Licensing Deal Yet
Nvidia licensed Poolside's Model Factory for $6B, invested $1B at a $12B valuation, and hired 109 staff to build a US open-weight challenger to China's best models.
-
Industry 中輝達豪擲 60 億美元買下 Poolside 的「模型工廠」——AI 史上最大授權交易
輝達以 60 億美元授權 Poolside 的 Model Factory、10 億美元投資（估值 120 億），並挖走 109 名工程師，打造抗衡中國開源模型的美國勁旅。
-
Tools ENTrueFoundry Open-Sources TrueForge: A Vendor-Neutral Agent Harness That Undercuts Claude Managed Agents by Up to 75%
MIT-licensed agent harness matches Claude Managed Agents' accuracy at 30-75% lower cost, and a $10,000 hackathon kicks off today.
-
Tools 中TrueFoundry 開源 TrueForge：中立廠商 Agent Harness，成本最多比 Claude Managed Agents 低 75%
MIT 授權的 agent harness 以相同準確率、最多便宜 75% 的姿態挑戰 Claude Managed Agents，萬美元黑客松今天登場。
-
Models ENThomson Reuters Launches 'Thomson': The $40M Legal LLM Betting It Can Beat the Frontier Labs
Thomson Reuters has launched Thomson, a legally trained LLM built on Alibaba's open-source Qwen — trained for $40M with $450K final runs, it already beats GPT-5.5 and Gemini 3.1 Pro on several legal benchmarks.
-
Models 中Thomson Reuters 推出自家 LLM「Thomson」：4000 萬美元訓練成本，要在法律領域擊敗前沿實驗室
Thomson Reuters 本週正式推出以阿里巴巴開源 Qwen 為基礎打造的法律專用 LLM「Thomson」——總投入約 4,000 萬美元、單次最終訓練僅 45 萬美元，卻已在多項法律基準測試擊敗 GPT-5.5 與 Gemini 3.1 Pro。
-
Industry ENAI Became AI's Biggest Customer: Agent Token Use Up 14x on OpenRouter Since February
OpenRouter data suggests February 6, 2026 was the last day humans out-consumed AI agents. Agent token usage has since grown 14x to 7.3 trillion tokens weekly — nearly 5x human levels — and cache economics are quietly rewriting the bill.
-
Industry 中AI 成了 AI 最大的客戶：OpenRouter 上代理權杖用量半年暴增 14 倍
OpenRouter 數據顯示，2026 年 2 月 6 日可能是人類最後一次在權杖消耗量上超過 AI 代理。半年來代理用量暴增 14 倍至每週 7.3 兆權杖——約人類的 5 倍——而快取經濟學正在悄悄改寫帳單。
-
Models ENAlibaba Officially Launches Wan3.0, a Document-to-Video AI Model, One Day After Its Record $10 Billion Share Sale
Alibaba Cloud's Wan3.0 turns decks, spreadsheets and web pages into 30-second videos — and lands just as the company raises a record HK$80 billion to fund its AI buildout.
-
Models 中阿里巴巴正式發布 Wan3.0 文件生影片模型，前一天才完成破紀錄 100 億美元增資
阿里雲 Wan3.0 能把簡報、試算表與網頁直接變成 30 秒影片——就在公司以破紀錄的 800 億港元增資挹注 AI 建設之際正式上線。
-
Industry ENReplaced by AI: Inside China's Growing Job Market Shock
From laid-off programmers to half-empty writer rooms, a new AP investigation shows how China's state-backed AI push is reshaping work faster than anywhere else on Earth.
-
Industry 中被 AI 取代：中國就業市場正在經歷的衝擊
從被裁員的程式設計師到剩下一半編劇的動畫公司，美聯社最新調查報導揭示了中國官方力推的 AI 浪潮如何比世界任何地方都更快地重塑勞動市場。
-
Industry ENAT&T Cut Its AI Bill 56% With a Router: The Quiet Rise of Enterprise 'Tokenomics'
By routing easy queries to Llama and Gemma, AT&T cut some coding-task costs 56% at a 2% quality cost — and now wants 60-70% of its 45 billion daily tokens on open models.
-
Industry 中AT&T 用一台路由器砍掉 56% AI 帳單：企業「Token 經濟學」的寧靜崛起
AT&T 透過把簡單查詢導向 Llama 與 Gemma，讓部分程式任務成本降低 56%、品質只掉 2%——下一步要讓每日 450 億 token 的六到七成跑在開源模型上。
-
Models ENDeepSeek's V4-Flash-Vision-Exp Edges Out Claude Opus 4.8 on Hard Vision Benchmarks
DeepSeek's experimental multimodal model adds image understanding at Flash-tier pricing, beating Claude Opus 4.8 on two hard visual benchmarks while staying radically cheaper.
-
Models 中DeepSeek V4-Flash-Vision-Exp 在高難度視覺基準測試中擊敗 Claude Opus 4.8
DeepSeek 的實驗性多模態模型以 Flash 級價格加入影像理解能力，在兩項高難度視覺基準測試中擊敗 Claude Opus 4.8，且價格便宜得多。
-
Industry ENThe Frontier Price War: OpenAI and Anthropic Slash Prices as Chinese Open-Weight Models Close In
OpenAI cut GPT-5.6 Luna by 80% and Anthropic locked Sonnet 5 at $2/$10 permanently — the fiercest frontier pricing battle yet, triggered by Chinese open-weight rivals like Kimi K3.
-
Industry 中前沿模型價格戰開打：OpenAI 與 Anthropic 大幅降價，中國開源開放權重模型步步進逼
OpenAI 將 GPT-5.6 Luna 降價 80%，Anthropic 把 Sonnet 5 永久鎖定在 $2/$10——這場由 Kimi K3 等中國開放權重模型引發的前沿定價大戰，是史上最激烈的一次。
-
Research ENAI4AI-Bench: The First Real Measurement of Recursive Self-Improvement Finds Agents Barely Off the Ground
A new benchmark asks LLM agents to rewrite the training algorithms that build AI itself. The best system closes under a fifth of the gap to optimal — and most agents never touch how the model learns at all.
-
Research 中AI4AI-Bench：首次實測「遞迴自我改進」，發現 AI 距離自我升級還很遠
新基準測試要求 LLM 代理改寫打造 AI 的訓練演算法本身。最強系統只走完到達最佳解不到五分之一的距離，而且多數代理根本沒碰模型的學習規則。
-
Tools EN"There's No Reason for Software to Be Slow Anymore": Dan Luu's Agent-Built Regex Engine and the Collapse of Performance Engineering Costs
A month-long agent loop built FRE, a regex engine that beats Rust's crate on long searches — and Dan Luu's follow-up experiments show performance work that once took specialist teams now takes minutes of human time.
-
Tools 中「軟體再也沒有理由變慢了」：Dan Luu 的代理自製 regex 引擎與效能工程成本的崩塌
一個跑了整月的代理迴圈打造出 FRE——在長查詢上擊敗 Rust regex crate 的引擎——而 Dan Luu 的後續實驗顯示：過去需要專家團隊的效能工程，如今只需幾分鐘的人力。
-
Models ENGLM-5.3: The Open Coding Model That Found 2,436 Real Bugs — and Its Own Weights Delayed
Z.ai's GLM-5.3 gets 50% better at coding from post-training alone, tops the CyberGym vulnerability benchmark at 84.5, and surfaced 2,436 real open-source vulnerabilities — so Z.ai delayed its own open-weights release for safety review.
-
Models 中GLM-5.3:找出 2,436 個真實漏洞的開源編程模型——連自己的權重都被延後發布
Z.ai 的 GLM-5.3 僅靠後訓練就讓編程能力提升 50%,在 CyberGym 漏洞挖掘基準以 84.5 奪冠,並找出 2,436 個真實開源漏洞——Z.ai 因此延後了自家開源權重的發布以進行安全審查。
-
Models ENOx Alpha: The Anonymous Frontier Model That Blindsided the AI World
A nameless 'stealth' model called Ox Alpha appeared on OpenRouter with a million-token context window, frontier-tier benchmarks, and a free week of near-unlimited access. Nobody knows who built it.
-
Models 中Ox Alpha：一個匿名前沿模型，讓整個 AI 圈瞬間沸騰
一個名為 Ox Alpha 的「隱身」模型悄悄現身 OpenRouter：百萬 token 上下文、前沿級基準成績、將近一週的免費暢用。沒有人知道它是誰做的。
-
Models ENA 27B Open-Weights Model Just Reverse-Engineered a Commercial App's License Check — Fully Offline
Qwen 3.8 27B, running entirely offline on a 128GB workstation, deconstructed a commercial app's licensing scheme, recovered an obscured crypto key, self-corrected its own mistake, and built a working bypass in 30 minutes.
-
Models 中270 億參數開源模型完全離線逆向商業軟體授權機制，30 分鐘完成
Qwen 3.8 27B 在一台 128GB 工作站上完全離線運行，拆解商業軟體的授權驗證、還原被混淆的加密金鑰、自我修正錯誤，並在 30 分鐘內做出可用的繞過概念驗證。
-
Tools ENLinus Torvalds Lets an AI Write a Linux Kernel Commit — After It Tried to Give Up
The Linux creator's rare personal patch fixes an Intel Xe driver bug behind a 24-patch, 18-boot debug marathon — and the commit message itself was written by AI.
-
Tools 中Linus Torvalds 讓 AI 寫下 Linux 核心提交訊息——在它多次想放棄之後
Linux 之父罕見親自出手修復 Intel Xe 驅動程式 bug,歷經 24 個偵錯補丁與 18 次開機——而提交訊息本身,是由 AI 寫的。
-
Industry ENHugging Face Is Exploring a $13 Billion Sale — and the Whole AI Stack Suddenly Has a Price Tag
Business Insider reports the open-source AI hub is working with a bank to field acquisition interest at $13B+ — nearly triple its 2023 valuation. Who buys the neutral ground of the AI economy?
-
Industry 中Hugging Face 傳探索 130 億美元出售——AI 開源生態的「中立地基」突然被標了價
Business Insider 獨家報導：開源 AI 平台 Hugging Face 正與銀行合作評估收購意向，估值至少 130 億美元，是 2023 年的三倍。誰能買下 AI 經濟的中立地帶？
-
Models ENTencent Quietly Slipped a Translation Specialist Onto OpenRouter — and It Undercuts Frontier APIs by 100x
Tencent Hunyuan's Hy-MT2 translation models landed on OpenRouter with no announcement: 33 language pairs, prices from $0.044/M tokens, and benchmark wins over DeepSeek-V4-Pro and Kimi K2.6.
-
Models 中騰訊低調把翻譯專用模型送上 OpenRouter——價格只有前沿 API 的百分之一
騰訊混元 Hy-MT2 翻譯模型家族無預警登上 OpenRouter：支援 33 種語言對、每百萬 token 低至 $0.044，並在基準測試擊敗 DeepSeek-V4-Pro 與 Kimi K2.6。
-
Research EN153 Runs, 18 Models, 8 Days Each: Prime Intellect Measured Whether AI Can Do Real Research
Prime Intellect pointed 18 frontier models at the nanoGPT speedrun and let them run unsupervised for up to 8 days. Fable 5 closed 81.7% of the human record gap — and not a single model invented a new method.
-
Research 中153 次自主運行、18 個前沿模型、每次最長 8 天：Prime Intellect 實測 AI 能不能做真正的研究
Prime Intellect 讓 18 個前沿模型在無人監督下挑戰 nanoGPT speedrun，最長連跑 8 天。Fable 5 收斂了人類紀錄差距的 81.7%——但沒有任何一個模型發明出 fundamentally 新的方法。
-
Tools ENZ.ai Launches OpenVuln: An AI Bug Hunter With a Public Paper Trail
Z.ai pairs its GLM-5.3 model with OpenVuln, a repo scanner that surfaced 2,436 real vulnerabilities — and a public ledger tracking every one to a fix.
-
Tools 中Z.ai 推出 OpenVuln：附公共揭露帳本的 AI 抓漏掃描器
Z.ai 為 GLM-5.3 配上程式碼庫掃描服務 OpenVuln，實測找出 2,436 個真實漏洞，並以公開帳本追蹤每一筆發現直到修復。
-
Tools ENAnthropic Launches Claude Academy: Free Courses, Badges, and a Skilljar Migration
Anthropic has quietly launched Claude Academy at academy.claude.com — a free learning hub with structured courses, quiz-earned completion badges, and Claude-account sign-in that replaces the old Skilljar platform for most learners.
-
Tools 中Anthropic 推出 Claude Academy：免費課程、完課徽章與 Skilljar 大搬遷
Anthropic 低調上線 Claude Academy（academy.claude.com）——免費學習平台，提供結構化課程、通過測驗即可獲得的完課徽章，並以 Claude 帳號登入取代多數學習者原本使用的 Skilljar。
-
Industry ENThe Best AI Model Nobody Buys: Ramp Data Shows Fable 5 Hit an Enterprise Spending Wall
Ramp's August AI Index shows Anthropic's flagship Fable 5 stuck at ~11% of business AI spend while the cheaper Opus 5 overtook it — the clearest signal yet that enterprises have found their price ceiling.
-
Industry 中沒人買的最強模型：Ramp 數據顯示 Fable 5 撞上企業支出高牆
Ramp 八月 AI 指數顯示，Anthropic 旗艦 Fable 5 在企業 AI 支出佔比停滯於約 11%，已被更便宜的 Opus 5 超越——這是企業買方找到價格天花板的最新明證。
-
Industry ENAlibaba Raises $10.2 Billion in Record Hong Kong Share Sale — Every Dollar Earmarked for AI
Alibaba launched a HK$80 billion ($10.2B) Hong Kong share placement on Sunday, pledging 100% of net proceeds to its full-stack AI buildout — the latest escalation in the global AI capex race.
-
Industry 中阿里巴巴香港史上最大規模配股募資 102 億美元，全數投入 AI
阿里巴巴週日在香港啟動 800 億港元（102 億美元）配股，承諾淨所得 100% 投入全端 AI 布局——這是全球 AI 資本競賽的最新升級。
-
Policy ENClaude Now Writes With an Invisible Watermark — Here's How the SynthID-Style Scheme Actually Works
Anthropic is embedding imperceptible watermarks in all Claude-generated text worldwide to comply with the EU AI Act, using a Google DeepMind-derived technique that bends randomness instead of words.
-
Policy 中Claude 開始在文字裡寫下看不見的浮水印——SynthID 式機制的真實原理
Anthropic 為符合歐盟 AI 法案，在全球範圍為所有 Claude 生成的文字嵌入無形浮水印——技術源自 Google DeepMind，彎曲的是隨機性而非文字本身。
-
Models ENGLM-5.3's Hacking Skills Outgrew Z.ai's Expectations — and Open Weights Are on Hold
Z.ai's new coding model found 2,436 real vulnerabilities in production software and became the first GLM release to hold back open weights for safety review.
-
Models 中GLM-5.3 的駭客能力超出 Z.ai 預期——開源權重因此暫緩發布
Z.ai 的新編程模型在正式環境軟體中找到 2,436 個真實漏洞,並成為首個因安全審查而暫緩開放權重的 GLM 版本。
-
Industry ENAlibaba-Backed Dexmal Targets $3 Billion Valuation in Embodied AI Funding Push
Chinese embodied-AI startup Dexmal is seeking a 20-billion-yuan ($3B) valuation in a new funding round, betting that warehouse picking — not chatbots — is the atomic task that unlocks general-purpose robots.
-
Industry 中阿里巴巴投資的 Dexmal 募資衝刺：瞄準 30 億美元估值，搶攻具身智慧
中國具身智慧新創 Dexmal 正在新一輪募資中尋求約 200 億人民幣（30 億美元）估值，押注倉儲揀選——而非聊天機器人——才是解鎖通用機器人的原子級任務。
-
Models ENAlibaba's Qwen Hits 3 Billion Downloads and Claims the Open AI Crown
Qwen has become the world's most downloaded open-weight AI model family, overtaking Meta and Google combined in just six months.
-
-
Tools ENThe Retrieval Layer Beat the Models: Pinecone Nexus Takes Top Score on τ-Knowledge
Same frontier models, different knowledge layer, better result: Pinecone Nexus hit GA and took the top score on Sierra's τ-Knowledge benchmark, beating agents built on OpenAI, Anthropic and Google.
-
Tools 中檢索層擊敗了模型：Pinecone Nexus 在 τ-Knowledge 基準測試奪下最高分
同樣的前沿模型、不同的知識層、更好的成績：Pinecone Nexus 正式版上市後，在 Sierra 的 τ-Knowledge 企業知識基準測試拿下最高分，擊敗了基於 OpenAI、Anthropic 與 Google 前沿模型打造的代理。
-
Industry ENKimi Splits in Two: Moonshot Separates General and Coding Memberships as Demand Overwhelms Compute
Moonshot AI is splitting Kimi subscriptions into separate general and coding tiers to ration scarce GPU capacity — the clearest sign yet that open-weight success has a compute bill attached.
-
Industry 中Kimi 一分為二：Moonshot 將一般與程式會員拆成雙軌制，需求壓垮算力
Moonshot AI 將 Kimi 訂閱拆成一般與程式兩種獨立會員，以分配稀缺的 GPU 算力——這是開源權重模型的成功同樣伴隨龐大運算帳單的最明確訊號。
-
Research ENZEST: Atlas Learns to Breakdance, Army-Crawl, and Backflip From Any Motion Source — Zero-Shot
Science Robotics' August humanoid special issue leads with ZEST, a motion-imitation framework from the RAI Institute and Boston Dynamics that trains whole-body policies from mocap, monocular video, or raw animation — then deploys them to Atlas, Unitree G1, and Spot with no per-skill engineering.
-
Research 中ZEST：Atlas 從任何動作來源學會地板舞、低爬與連續後空翻——零樣本上機
《Science Robotics》八月人形機器人專刊以 ZEST 為封面論文：這套由 RAI Institute 與 Boston Dynamics 開發的動作模仿框架，能從動作捕捉、單眼影片或純動畫訓練全身控制策略，並零樣本部署到 Atlas、Unitree G1 與 Spot 上，無需逐技能工程。
-
Tools ENMCP's Next Act: Agentic Messaging, Agent Identity, and One HTTP Transport
The Model Context Protocol's lead maintainers published a new roadmap on August 22, 2026, reorienting the spec around agent-native messaging, unified HTTP transport, and standardized agent identity — the plumbing the agentic web will run on.
-
Tools 中MCP 的下一步：代理式訊息傳遞、Agent 身分識別，與單一 HTTP 傳輸層
Model Context Protocol 首席維護者於 2026 年 8 月 22 日發布全新路線圖，將規格重心轉向代理原生訊息傳遞、統一 HTTP 傳輸與標準化 Agent 身分——這是代理式網路未來賴以運作的基礎管線。
-
Industry ENMicron Bets $10 Billion on a Decade of Post-DRAM Research in Boise
Micron Research Labs will chase memory, compute, and packaging breakthroughs a decade out — the first US hub of its kind, backed by everyone from NVIDIA to Stanford.
-
Models ENOx Alpha: The Mystery Frontier Model Beating GPT-5.6 at Coding — and It's Free
An anonymous stealth model dubbed Ox Alpha appeared on OpenRouter this week with a 1M-token context window, 100 trillion free tokens per day, and coding scores that top GPT-5.6 — and fingerprinting evidence points straight at Zhipu's unreleased GLM-5.x.
-
Models 中Ox Alpha：擊敗 GPT-5.6 的神秘前沿模型——而且免費
一個名為 Ox Alpha 的匿名隱身模型本週現身 OpenRouter，配備百萬 token 上下文視窗、每日 100 兆免費 token，程式編寫評測超越 GPT-5.6——種種指紋證據直指智譜未發布的 GLM-5.x。
-
Models ENDeepSeek Gives Its Cheapest Model Eyes: V4-Flash-Vision-Exp Lands Within Striking Distance of Opus-4.8
DeepSeek's experimental multimodal model adds image understanding to its 284B-parameter MoE at V4-Flash prices, beating Anthropic's Opus-4.8 on three of eleven agent benchmarks and matching it on several more.
-
Models 中DeepSeek 給最便宜的王牌裝上眼睛：V4-Flash-Vision-Exp 逼近 Opus-4.8
DeepSeek 的實驗性多模態模型以 V4-Flash 的價格為 284B 參數 MoE 架構加入影像理解，在十一項代理基準中三項擊敗 Anthropic 的 Opus-4.8，其餘多項僅以些微差距落後。
-
Meta ENAnthropic Puts Claude Mythos 5 to Work: Cyber Supermodel Now Scans Enterprise Code, Backed by a $35M Open-Source Defense Fund
Anthropic has deployed Claude Mythos 5 — its most cyber-capable model, restricted to vetted defenders since April — inside Claude Security for all Enterprise customers, launched a $35M Defender Advantage Fund for open-source protection, and previewed a wider Cyber Verification Program.
-
Meta 中Anthropic 讓 Claude Mythos 5 上工：網安超級模型開始掃描企業程式碼，同步成立 3,500 萬美元開源防禦基金
Anthropic 將自 4 月起僅限審核通過防禦者使用的網安旗艦模型 Claude Mythos 5，部署到 Claude Security 供所有企業客戶掃描漏洞，並成立 3,500 萬美元的 Defender Advantage Fund 協助保護開源軟體。
-
Research ENSemiAnalysis: Open Models Now Close the Frontier Gap in Half the Time Every AI Era
A new SemiAnalysis study finds open-weight models are matching closed frontier models twice as fast with each successive AI era — from 12 months in the scaling era to under 5 months today, while Fireworks alone now serves 40 trillion tokens a day.
-
Research 中SemiAnalysis 研究揭露:開源模型追上閉源前沿的速度,每個 AI 時代都縮短一半
SemiAnalysis 最新研究發現,開源權重模型追平閉源前沿模型所需的時間,隨每個 AI 時代以減半速度壓縮——從擴展時代的 12 個月,到如今不到 5 個月,而 Fireworks 單日處理量已突破 40 兆 token。
-
Models ENGeneralist's GEN-1.5 Turns Robots Into One-Shot Learners From a Single 3-Second Demo
Robotics startup Generalist says its GEN-1.5 foundation model learns new physical tasks from one 3–12 second demonstration with no training at all — 59% success zero-shot, 83% after ten gradient steps.
-
Models 中Generalist 推出 GEN-1.5：機器人從單次 3 秒示範學會新任務
機器人新創 Generalist 發表 GEN-1.5 基礎模型：僅憑一段 3–12 秒的示範就能學會新的物理任務，完全無需訓練——零梯度更新成功率 59%，十步梯度微調後達 83%。
-
Models ENGemma Passes 1 Billion Downloads: Google's Open Weights Now Run From Orbit to the Ocean Floor
Google DeepMind says its Gemma open models passed 1 billion cumulative downloads, with 100,000+ community variants running everywhere from satellites in orbit to India's national health app.
-
Models 中Gemma 下載量突破 10 億次:Google 的開放權重模型已從軌道部署到深海
Google DeepMind 宣布 Gemma 開放模型累積下載量突破 10 億次,社群變體超過 10 萬個,部署範圍從軌道上的衛星到印度的全國健康應用。
-
Industry ENNvidia Pays Poolside $6 Billion to License Its 'Model Factory' — the Third Mega-Deal in Its New Playbook
Nvidia licensed Poolside's Model Factory for $6B, hired 109 staff, and invested $1B at a $12B valuation — without buying the company. It's the third deal built on this exact template.
-
Industry 中Nvidia 斥資 60 億美元取得 Poolside「模型工廠」授權——新版交易模板的第三案
Nvidia 以 60 億美元授權 Poolside 的 Model Factory、延攬 109 名員工，並以 120 億美元估值投資 10 億美元——卻沒有買下這家公司。這已是同一套劇本的第三次演出。
-
Models ENGemma Hits One Billion Downloads: Inside Google's Open-Weights Gambit
Google DeepMind's Gemma family has passed one billion downloads with over 100,000 community variants, and Google is consolidating the 'Gemmaverse' into an official Awesome Gemma hub — the open-weights race now has a scoreboard.
-
Models 中Gemma 下載量突破十億：解析 Google 的開放權重戰略
Google DeepMind 的 Gemma 系列開放模型累積下載量突破 10 億次、社群變體超過 10 萬個，並推出官方 Awesome Gemma 目錄——開放權重之戰正式有了計分板。
-
Models ENGLM-5.3: Z.ai's Cyber-Surprise Model Is So Good at Hacking It's Holding Its Own Weights
Z.ai's GLM-5.3 matches frontier coders on 743B parameters — but emergent exploit skills it never trained for forced a two-week delay of the open weights.
-
Models 中GLM-5.3：強到會自己找漏洞的開源模型，Z.ai 被迫扣住自家權重不發
Z.ai 的 GLM-5.3 以 743B 參數追平前線閉源編碼模型，但測試中「意外長出」的攻擊能力，讓開源權重被迫延後兩週發布。
-
Models ENDeepSeek V4-Flash-Vision-Exp: The Experimental Multimodal Model Chasing Opus 4.8
DeepSeek's experimental vision-equipped V4-Flash lands within points of Anthropic's Opus 4.8 on multimodal agent benchmarks — at a fraction of the price.
-
Models 中DeepSeek V4-Flash-Vision-Exp：追趕 Opus 4.8 的實驗性多模態模型
DeepSeek 的實驗性視覺版 V4-Flash 在多模態代理基準上逼近 Anthropic Opus 4.8，價格卻只有零頭。
-
Policy ENBrazil Splits $444M AI Supercomputer Bet Between US and China
Brazil announced 2.3 billion reais in AI infrastructure — an Nvidia-expected national supercomputer in Rio Grande do Norte and a Huawei-iFlytek LLM cluster in Rio — hedging between both superpowers.
-
Policy 中巴西 4.44 億美元 AI 豪賭：超級電腦一半給美國、一半給中國
巴西宣布投入 23 億里爾建設 AI 基礎設施——北里奧格蘭德的美系國家超級電腦加上里約的華為・科大訊飛 LLM 叢集——在兩強之間兩邊下注。
-
Industry ENClaude's Watermark Backlash: Open-Source Removers Hit 12,000 GitHub Stars as Users Cancel Subscriptions
Anthropic's invisible Claude text watermarks were meant to satisfy the EU AI Act — instead they triggered user cancellations and an open-source removal arms race that hit 12,000 GitHub stars in under two weeks.
-
Industry 中Claude 浮水印引發反彈：開源移除工具兩週破萬星，用戶退款取消訂閱
Anthropic 為符合 EU AI Act 而在 Claude 文字輸出嵌入隱形浮水印，沒想到引發訂戶取消潮，並催生出兩週內衝上 12,000 星的開源移除工具軍備競賽。
-
Research ENLinear's Data Shows AI Now Writes Half of All Issues — and Teams Are Working More, Not Less
Linear's first 'How Teams Build' report finds agents author ~49% of all issues, coding-agent teams tripled weekly PRs, and total time spent on product development is rising — a Jevons paradox for the AI era.
-
Research 中Linear 數據報告：AI 已寫下近半數 Issue——但團隊工時不減反增
Linear 首份《How Teams Build》報告顯示：Agent 與 MCP 客戶端已撰寫約 49% 的 issue，接上編碼代理的團隊每週 PR 數翻三倍，但產品開發總工時持續上升——AI 時代的 Jevons 悖論。
-
Industry ENNvidia Pays $6 Billion to License Poolside's AI Models and Hires Its Team
Nvidia agreed to pay $6 billion for a non-exclusive license to Poolside's Model Factory code-generation models, invested $1 billion at a $12 billion pre-money valuation, and extended offers to all 109 Poolside employees — the chipmaker's latest mega-scale acqui-hire.
-
Industry 中Nvidia 斥資 60 億美元取得 Poolside AI 模型授權，並把整個團隊挖進門
Nvidia 同意支付 60 億美元取得 Poolside Model Factory 程式碼模型的非專屬授權，另以 120 億美元投前估值投資 10 億美元，並向全部 109 名員工提出轉職邀約——這是晶片巨頭最新、也最大規模的超級授權加人才收購案。
-
Research ENOpenBMB Open-Sources Ultra-FineWeb-L1: 1.3T Tokens of 2025 Web Data Under Apache 2.0
OpenBMB releases Ultra-FineWeb-L1, a 1.3-trillion-token English web corpus built from six 2025 Common Crawl snapshots — the freshest open pretraining dataset to date, beating FineWeb by 0.6 points in ablation runs.
-
Research 中OpenBMB 開源 Ultra-FineWeb-L1：1.3 兆 Token 的 2025 年網頁語料，Apache 2.0 授權釋出
OpenBMB 發布 Ultra-FineWeb-L1——以六個 2025 年 Common Crawl 快照構建的 1.3 兆 token 英文網頁語料庫，是迄今涵蓋最新爬取快照的開源預訓練資料集，消融實驗中較 FineWeb 高出 0.6 分。
-
Industry ENCallosum Raises $100M Seed to Route AI Workloads Across Mixed Chips — With the British State on the Cap Table
Cambridge spinout Callosum landed one of Europe's largest-ever seed rounds — $100M led by Atomico, with the UK's Sovereign AI Unit on the cap table — to build chip-agnostic orchestration software that claims 4x speed and 70% cost savings on agentic workloads.
-
Industry 中Callosum 完成 1 億美元種子輪：讓 AI 工作負載跨晶片調度，英國政府直接入股
劍橋大學新創 Callosum 拿下歐洲史上最大種子輪之一——由 Atomico 領投的 1 億美元，英國主權 AI 基金也名列股東——其晶片無關的編排軟體宣稱在代理式工作負載上可帶來 4 倍速度與 70% 成本節省。
-
Industry ENAlibaba's AI Gamble: Profit Crashes 75% as Cloud Revenue Rockets 45%
Alibaba's June quarter shows the brutal economics of the AI buildout: cloud revenue up 45%, profit down 75%, capex up 75% to $10B a quarter — and the market cheered anyway.
-
Industry 中阿里巴巴的 AI 豪賭：雲端營收暴增 45%，獲利卻崩跌 75%
阿里巴巴 2026 年 6 月季度財報揭示了 AI 建軍的殘酷經濟學：雲端營收年增 45%、獲利大減 75%、單季資本支出暴增 75% 至 100 億美元——而市場竟然拍手叫好。
-
Tools ENSnowflake's Cortex AI Gateway Now Picks Your Model for You — and Cuts Token Spend Up to 3x
Snowflake's dynamic model routing auto-selects the cheapest model that clears the quality bar, adding DeepSeek-V4-Flash and GLM-5.3 while keeping data governed.
-
Tools 中Snowflake Cortex AI Gateway 學會自己挑模型——token 開銷最高省 3 倍
Snowflake 推出動態模型路由,自動為每個請求挑選「夠用且最省」的模型,同時納入 DeepSeek-V4-Flash 與 GLM-5.3 開源模型,資料全程留在治理邊界內。
-
Research ENClaude Designed Working Protein Binders for 14 of 15 Targets — No Human in the Loop
Anthropic ran Claude autonomously through full protein binder design campaigns: 354 of 1,320 designs bound in wet-lab tests, beating open competition entries on hit rate and affinity.
-
Research 中Claude 自主設計蛋白質結合器，15 個目標中 14 個成功——全程無人干預
Anthropic 讓 Claude 全自主執行完整蛋白質結合器設計流程：1,320 個設計中有 354 個通過濕實驗室驗證，命中率和親和力均超越公開競賽作品。
-
Tools ENReplit Free Mode: 30x More Building for $20 a Month, Powered by GPT-5.6 Luna
Replit's new Free Mode, built with OpenAI's cost-efficient GPT-5.6 Luna, lets Core subscribers create up to 30x more without burning credits — a new phase in the AI app-builder price war.
-
Tools 中Replit 推出 Free Mode:每月 20 美元創作量提升 30 倍,背後是 GPT-5.6 Luna
Replit 攜手 OpenAI 推出 Free Mode,以高效能比的 GPT-5.6 Luna 為核心,讓訂閱戶日常創作不再消耗點數,打響 AI 應用開發平台的降價戰。
-
Tools ENVercel Open-Sources fx: A 6.39 MiB Zig Coding Agent That Cold-Starts in 10 Microseconds
Vercel Labs' fx is a minimalist coding agent harness written in Zig — a 6.39 MiB binary with 10µs cold starts, Wasm builds, Apache-2.0 license, and a Unix-shell philosophy that challenges the bloated coding-agent status quo.
-
Tools 中Vercel 開源 fx：以 Zig 打造、僅 6.39 MiB 且冷啟動 10 微秒的編碼代理
Vercel Labs 的 fx 是以 Zig 撰寫的極簡編碼代理框架——6.39 MiB 的二進位檔、10 微秒冷啟動、支援 Wasm、採 Apache-2.0 授權，以 Unix 哲學挑戰日益臃腫的編碼代理生態。
-
Meta ENCerebras CS-4: Wafer-Scale Inference Grows Up Into a Rack
Cerebras unveils the CS-4, a rack-scale system built on three WSE-3 Turbo wafers claiming up to 30x faster inference than GPUs and over 1,000 tokens/sec on 10T-parameter models.
-
Meta 中Cerebras CS-4：晶圓級推論正式長成一整個機櫃
Cerebras 發表 CS-4，以三片 WSE-3 Turbo 晶圓打造的機櫃級系統，宣稱推論速度最快可達 GPU 的 30 倍，10 兆參數模型每秒可產出超過 1,000 個 token。
-
Industry ENStripe Acquires OpenRouter: Tokens Become the New Currency of the AI Economy
Stripe has agreed to acquire AI model gateway OpenRouter for over $7 billion, betting that intelligent token routing will become the economic infrastructure of the AI era.
-
Industry 中Stripe 收購 OpenRouter：Token 成為 AI 經濟的新貨幣
Stripe 宣布以超過 70 億美元收購 AI 模型閘道 OpenRouter，押注智慧型 token 路由將成為 AI 時代的經濟基礎設施。
-
Tools ENHiggsfield's 'The Cully Hill Boys': The First Fully AI-Generated Feature Film Goes Mainstream
Higgsfield AI's 110-minute action-comedy 'The Cully Hill Boys' — built in four weeks for $2M with licensed celebrity likenesses — is drawing national coverage as the moment AI cinema crossed from demo to deliverable.
-
Tools 中Higgsfield《The Cully Hill Boys》：首部全 AI 生成的長片電影走向主流
Higgsfield AI 的 110 分鐘動作喜劇《The Cully Hill Boys》以 200 萬美元預算、四週工期與授權名人肖像打造，隨著 Semafor 於 8 月 19 日發表評論，AI 電影正式從技術演示跨入可交付的商業製品。
-
Policy ENChina Pushes Back on US 'Pick Sides' AI Ultimatum, Urging Digital Sovereignty for All
Beijing opposes bloc-building in AI after a draft US letter told 35 countries to choose between the American ecosystem and China's WAICO framework.
-
-
Industry ENRillet Raises $100M at $1B Valuation to Put AI Agents Inside the General Ledger
Rillet's ICONIQ-led Series C values the AI-native ERP at $1B after new ARR doubled in three months — and its bet that finance agents belong inside the general ledger, not bolted on top, is becoming the template for agentic enterprise software.
-
Industry 中Rillet 以 10 億美元估值完成 1 億美元 C 輪融資，讓 AI 代理人直接走進總帳
Rillet 獲 ICONIQ 領投的 C 輪融資、估值達 10 億美元，新簽年經常性收入三個月內翻倍。該公司押注金融代理人應該「在總帳裡面」工作而非外掛在其上，這套架構正逐漸成為代理式企業軟體的新範本。
-
Research ENClaude Designs Working Protein Binders Against 14 of 15 Targets: Anthropic's Lab-Validated Biology Results
Anthropic reports Claude autonomously designed de novo protein binders that succeeded against 14 of 15 wet-lab targets with 22–35% hit rates — more than double the field's 10–15% norm — plus NMR/LC-MS analysis in minutes instead of days.
-
Research 中Claude 設計出 15 個標的中 14 個可用的蛋白質結合子：Anthropic 的實驗室驗證生物學成果
Anthropic 發布濕實驗室驗證結果：Claude 在 Claude Science 環境中自主設計全新蛋白質結合子，15 個標的中成功命中 14 個，命中率 22–35%——是業界常態 10–15% 的兩倍以上，並將 NMR/LC-MS 分析從數天縮短到 23 分鐘。
-
Policy ENBeijing Fires Back at Washington's AI Ultimatum: 'Each Country Has the Right to Choose Its Partners'
Five days after Reuters revealed a draft US letter telling 35 countries to choose between Pax Silica and China's WAICO, Beijing went on the record: no forced side-taking, no AI camps — respect digital sovereignty instead.
-
Policy 中北京正式回擊華府 AI 最後通牒：「各國有權自主選擇合作夥伴」
路透社披露美國草擬信函、要求 35 國在 Pax Silica 與中國 WAICO 之間二選一的五天後，北京外交部首次公開表態：反對選邊站隊、反對陣營對抗，呼籲尊重數位主權。
-
Industry ENTemporal in Talks to Raise at a $12 Billion Valuation as Agentic AI Demand Explodes
Bloomberg reports Temporal Technologies is negotiating a fresh round of roughly $500 million at a valuation of at least $12 billion, more than doubling its February price tag.
-
Industry 中Temporal 傳洽談以至少 120 億美元估值募資：Agentic AI 需求爆發下的基礎設施贏家
Bloomberg 報導，開源持久化執行平台 Temporal 正洽談募集約 5 億美元新一輪資金，估值至少 120 億美元，較今年 2 月的 50 億美元翻倍有餘。
-
Tools ENBlock Open-Sources Berd: A Local-First Desktop Workspace for AI Agents Built on Goose
Block has released Berd, the Tauri 2 desktop app its own teams use to run AI agents across projects, skills, and models, under Apache 2.0 — a local-first, model-agnostic alternative to browser-based agent workspaces.
-
Tools 中Block 開源 Berd：基於 Goose 的本地優先 AI Agent 桌面工作區
Block 將內部團隊日常使用的 AI agent 桌面應用 Berd 以 Apache 2.0 授權開源——以 Tauri 2 打造、透過 ACP 協定對接 Goose 後端，主打本地優先、模型中立，成為瀏覽器式 agent 工作區之外的另一種選擇。
-
Models ENCartesia's Sonic-3.6 Seizes #1 on Both Artificial Analysis Speech Arenas
Cartesia ships Sonic-3.6, a state-space streaming TTS model that now tops both Artificial Analysis speech leaderboards — 1,283 Elo on Provider Voice and #1 on the Controlled Voice board that isolates the synthesis engine itself.
-
Models 中Cartesia Sonic-3.6 登上 Artificial Analysis 兩大語音競技場冠軍
Cartesia 推出 Sonic-3.6 串流語音合成模型,採用狀態空間架構,同時拿下 Artificial Analysis 兩大語音排行榜第一 —— Provider Voice 以 1,283 Elo 稱王,Controlled Voice 排行榜更是驗證了合成引擎本身的實力。
-
Tools ENThree Days to Patch: CISA Orders Emergency Fix for Actively Exploited Ray AI Framework Flaw
CISA has given US federal agencies until August 20 to patch CVE-2025-62593, a critical RCE in the Ray AI compute framework that turns a developer's Firefox or Safari browser into a weapon against corporate ML clusters.
-
Tools 中三天內修補：CISA 緊急命令修復遭活用攻擊的 Ray AI 開源框架漏洞
CISA 要求美國聯邦機構在 8 月 20 日前修補 CVE-2025-62593——Ray AI 運算框架中的重大遠端程式碼執行漏洞，攻擊者能把開發者的 Firefox 或 Safari 瀏覽器變成打入企業 ML 叢集的武器。
-
Tools ENWarp Factories: The Out-of-the-Box Software Factory for AI-Native Engineering Teams
Warp launched Warp Factories on August 18 — a ready-made infrastructure layer that runs AI agents across triage, spec, implementation, review and verification, with bring-your-own models, Linear/Jira/Slack integration, token-spend analytics and self-improvement loops for teams that can't build a Stripe-scale 'minions' system themselves.
-
Tools 中Warp Factories：為 AI 原生工程團隊而生的「開箱即用軟體工廠」
Warp 於 8 月 18 日推出 Warp Factories——一個現成的基礎架構層，讓 AI agent 負責分類、規格、實作、審查與驗證，支援自帶模型、整合 Linear/Jira/Slack、提供 token 消費分析與自我改善迴圈，專為無法自建 Stripe 級「minions」系統的團隊而設計。
-
Policy ENHollywood's First AI Copyright Truce: MPA and ByteDance Sign Landmark Guardrail Pact
The Motion Picture Association and ByteDance signed the first-ever copyright agreement between Hollywood and an AI company — covering Seedance, Seedream, TikTok, and CapCut, six months after a viral deepfake nearly triggered litigation.
-
Models ENGLM-5.3: The Open-Weight Model That Got Scary Good at Coding — and Found 2,436 Real Vulnerabilities
Z.ai's GLM-5.3 uses the exact same base model as GLM-5.2 — every gain came from post-training. It jumped from 4.6 to 28.3 on Terminal-Bench 3.0, topped the CyberGym security benchmark at 84.5%, and surfaced 2,436 real vulnerabilities in production code, some 40 years old. Weights land in two weeks.
-
Models 中GLM-5.3：同一個基座模型，後訓練就讓它 coding 強到嚇人——還找出 2,436 個真實漏洞
Z.ai 的 GLM-5.3 與 GLM-5.2 用的是完全相同的基座模型——所有進步都來自後訓練。Terminal-Bench 3.0 從 4.6 跳到 28.3，以 84.5% 登頂 CyberGym 安全基準，並在生產程式碼中挖出 2,436 個漏洞、最老的已潛伏約 40 年。權重兩週後開放。
-
Models ENQwen3.8-27B Lands on Laptops as Alibaba Unlocks Qwen3.8-Max: The Open-Weight War Goes Two-Front
Alibaba answered Meta's Muse Glimmer within a week: a laptop-ready Qwen3.8-27B with Opus-class coding benchmarks, plus free downloads of its 2.4-trillion-parameter flagship. The open-weight race now runs from datacenters down to your MacBook.
-
Models 中Qwen3.8-27B 登上筆電、Qwen3.8-Max 開放權重：阿里巴巴的開源雙線戰爭
Meta 推出 Muse Glimmer 不到一週，阿里巴巴隨即雙管齊下：釋出可在筆電上運行、編碼能力媲美 Opus 的 Qwen3.8-27B，同時免費開放 2.4 兆參數旗艦模型 Qwen3.8-Max 的完整權重。開源模型之戰，從資料中心一路燒到你桌上的 MacBook。
-
Models ENGemini 3.7 Flash: Google's Coding Workhorse Gets a 50% Price Cut
Google ships Gemini 3.7 Flash just three weeks after 3.6 — DeepSWE coding score jumps from 49% to 65.3% at half the price.
-
Models 中Gemini 3.7 Flash：Google 編碼主力模型降價 50%
Google 在 3.6 Flash 發布僅三週後就推出 Gemini 3.7 Flash——DeepSWE 編碼評測從 49% 躍升至 65.3%，價格砍半。
-
Industry ENStripe Buys OpenRouter for $7B+: Payments Giant Owns the AI Model Gateway
Stripe finalized a deal to acquire OpenRouter for more than $7 billion — a 5x markup on the AI gateway's May valuation — betting that routing between 400+ AI models becomes the billing rail for the agent economy.
-
Industry 中Stripe 以超過 70 億美元收購 OpenRouter：支付巨頭親自拿下 AI 模型閘道
Stripe 敲定以逾 70 億美元收購 OpenRouter——較這家 AI 閘道公司五月估值翻漲逾五倍——押注在 400 多個模型之間路由，將成為代理經濟的帳務軌道。
-
Research ENThe Erdős Gold Rush: How a Dead Mathematician's Problem List Became AI's Toughest Benchmark
Over a hundred of Paul Erdős's open problems have fallen since October 2025 — to hobbyists with chatbots, Google DeepMind agents, and OpenAI's unreleased Astra model. Quanta's deep dive explains why this quirky list became the proving ground where AI learned to do real mathematics.
-
Research 中Erdős 問題淘金熱：一位已故數學家的問題清單，如何變成 AI 最嚴苛的基準測試
自 2025 年 10 月以來，Erdős 的上百道開放問題接連被攻克——解題者包括拿聊天機器人的業餘愛好者、Google DeepMind 的代理，以及 OpenAI 尚未發表的 Astra 模型。Quanta 的深度報導解釋了這份古怪清單為何成為 AI 學會「做真正的數學」的試煉場。
-
Models ENMeta's Muse Glimmer: A 30B Open-Weight Agent That Runs on a Single Gaming GPU
Meta returns to open weights with Muse Glimmer, a 30-billion-parameter Apache 2.0 model built for always-on local AI agents — no datacenter, no API bill, no rate limits.
-
Models 中Meta Muse Glimmer：30B 開放權重代理模型，一張電競顯卡就能離線跑
Meta 重返開放權重陣營：Muse Glimmer 是 300 億參數、Apache 2.0 授權、專為本機常駐 AI 代理打造的模型——不需要資料中心、不需要 API 帳單、沒有速率限制。
-
Industry ENQwen Hits 3 Billion Downloads: Alibaba's Open-Weight Empire and the Hugging Face Reality Check
Alibaba claims Qwen has passed 3 billion downloads to become the world's most-used open model family. Hugging Face's own summer report counts 2.05 billion — and shows why Qwen has become the community's default base model anyway.
-
Industry 中Qwen 下載量突破 30 億：阿里巴巴的開放權重帝國與 Hugging Face 的數據校正
阿里巴巴宣稱 Qwen 下載量突破 30 億，成為全球最多人使用的開放模型家族；Hugging Face 夏季報告自己的統計是 20.5 億——但無論用哪個數字，Qwen 都已成為開發者社群的預設基礎模型。
-
Industry ENDatabricks Closes $5 Billion Round at $190 Billion Valuation as Growth Accelerates Past 80%
Databricks sealed a $5 billion strategic round at a $190 billion valuation, surpassing a $7 billion revenue run-rate with >80% YoY growth — one of the largest private financings in software history.
-
Industry 中Databricks 以 1,900 億美元估值完成 50 億美元融資，營收增速突破 80%
Databricks 以 1,900 億美元估值完成 50 億美元策略性融資，營收年化 run-rate 突破 70 億美元、年增超過 80%——這是軟體史上規模最大的私人融資之一。
-
Tools ENGoogle Kills Imagen 4: Image API Dies Today, Nano Banana 2 Takes Over
Google's Imagen 4 API family — standard, Ultra, and Fast — reaches end-of-life on August 17, 2026. Here's what breaks, what it costs, and how to migrate to Gemini 3.1 Flash Image (Nano Banana 2).
-
Tools 中Google 終結 Imagen 4:圖像 API 今日走入歷史,Nano Banana 2 接棒
Google 的 Imagen 4 API 家族——標準版、Ultra 與 Fast——於 2026 年 8 月 17 日終止服務。哪些東西會壞掉、成本如何變化、又該如何遷移到 Gemini 3.1 Flash Image(Nano Banana 2),本文一次解析。
-
Industry ENDeepSeek Raises API Prices Up to 11x: The Price War Leader Flips the Board
The lab that detonated the AI price war now charges up to 11 times more for its V4 models, with peak-hour billing — a signal that below-cost inference is over across the industry.
-
Models ENGoogle's Gemini 3.7 Flash: The Half-Price Workhorse Built for the Agent Era
Three weeks after 3.6 Flash, Google ships Gemini 3.7 Flash with big coding gains, state-of-the-art agent benchmarks, and a 50% introductory price cut to $0.75 per million input tokens.
-
Models 中Google Gemini 3.7 Flash：為代理時代而生、價格砍半的工作馬模型
距離 3.6 Flash 僅三週，Google 推出 Gemini 3.7 Flash：程式碼能力大幅躍進、代理任務跑出 SOTA，輸入價格更砍至每百萬 token 僅 0.75 美元。
-
Models ENMeta's Muse Glimmer Puts a 30B Agentic Model on Your Laptop — Under Apache 2.0
Meta's first major open-weight release under AI chief Alexandr Wang is a 30-billion-parameter, vision-capable agent model that runs on a single consumer GPU — and it's free for commercial use.
-
Models 中Meta Muse Glimmer：30B 開源代理模型，一張消費級顯卡就能跑
Meta 在新任 AI 負責人 Alexandr Wang 麾下的首波重大開放權重發布：300 億參數、具備視覺能力的代理模型，單張消費級 GPU 即可運行，Apache 2.0 授權可商用。
-
Policy ENAnthropic's Claude Now Watermarks Everything It Writes
Claude models launched on or after August 2 now embed an invisible watermark in all generated text and sign files with C2PA provenance metadata — here's how it works and why the EU AI Act forced the change.
-
Policy 中Anthropic 的 Claude 開始為所有生成內容加上隱形浮水印
2026 年 8 月 2 日後推出的 Claude 模型，現在會在所有生成文字中嵌入隱形浮水印，並為檔案附加 C2PA 簽署來源 metadata——本文解析其運作原理，以及歐盟 AI 法案如何促成這項改變。
-
Models ENMotif 3 Ships Final Weights Under MIT License: Korea's Sovereign AI Goes Fully Open
Korea's Motif Technologies quietly released the final 314B-parameter Motif 3 under MIT license — a from-scratch MoE architecture that scores 47 on the Artificial Analysis Intelligence Index.
-
Models 中Motif 3 正式版以 MIT 授權開放：韓國主權 AI 全面開源
韓國 Motif Technologies 低調釋出 314B 參數的 Motif 3 正式版權重，採 MIT 授權——從零打造的 MoE 架構，在 Artificial Analysis 智慧指數拿下 47 分。
-
Models ENQwen3.8-27B Ships With Native Vision, 262K Context and Apache 2.0 Weights
Alibaba's compact 27B open-weight model lands with a built-in vision encoder, 262K native context and agentic coding scores that embarrass models twice its size.
-
Models 中Qwen3.8-27B 開源登場：原生視覺、262K 上下文、Apache 2.0 授權
阿里巴巴以 27B 精巧開源模型內建視覺編碼器與 262K 原生上下文，Agentic 編碼分數讓兩倍大的模型顏面無光。
-
Models ENGLM-5.3: Z.ai's Post-Training Experiment Unleashes Emergent Cyber Capabilities
Z.ai shipped GLM-5.3 on the same 743B base as GLM-5.2 — post-training alone doubled exploit benchmarks and produced unplanned offensive security skills, including a reported serious vulnerability in Cursor.
-
Models 中GLM-5.3：Z.ai 的後訓練實驗催生了非預期的網安能力
Z.ai 在與 GLM-5.2 完全相同的 743B 基礎模型上推出 GLM-5.3——純靠後訓練就讓漏洞利用基準翻倍，甚至產生了訓練計畫之外的攻擊能力，據報導已找到 Cursor 的嚴重漏洞。
-
Models ENNVIDIA's Nemotron 3.5 Lightning Is a 30B Open Model Built to Do the Agent Grunt Work
NVIDIA's new open 30B MoE model with just 3B active parameters targets the high-volume execution layer of always-on agents — 4x faster output, 54.3% on SWE-Bench Verified, and a 1M-token context, all under the permissive OpenMDW license.
-
Models 中NVIDIA Nemotron 3.5 Lightning：專門處理 Agent 苦差事的 30B 開源模型
NVIDIA 最新開源 30B MoE 模型僅啟用 3B 參數，專為常駐型 Agent 的高頻執行層而設計——輸出速度最快達 4 倍、SWE-Bench Verified 拿下 54.3%，支援 1M token 上下文，並採用寬鬆的 OpenMDW 授權。
-
Tools ENKog's Bet: 30x Faster LLM Inference Without Buying a Single New GPU
French startup Kog hit 3,000 tokens per second on stock AMD MI300X and Nvidia H200 GPUs with pure software optimization — and now it's racing to bring that speed to full-size LLMs by September.
-
Tools 中Kog 的豪賭：不買新 GPU，照樣讓 LLM 推理快 30 倍
法國新創 Kog 只靠軟體最佳化，就在標準的 AMD MI300X 與 Nvidia H200 上跑出每秒 3,000 tokens 的推理速度——現在正趕著在九月把這套技術推向完整尺寸的大型語言模型。
-
Industry ENQwen Hits 3 Billion Downloads: Alibaba's Open Models Now Outpace Meta and Google Combined
Alibaba's Qwen family has passed 3 billion downloads in six months — versus 418M for Google and 227M for Meta in all of 2026. Combined with Hugging Face's new State of Open Models report, the numbers confirm the open-weights center of gravity has shifted decisively to China.
-
Industry 中Qwen 下載量突破 30 億：阿里開源模型正式超越 Meta 與 Google 的總和
阿里巴巴的 Qwen 系列在六個月內累積超過 30 億次下載——而 Google 與 Meta 在 2026 全年分別僅約 4.18 億與 2.27 億。搭配 Hugging Face 最新發布的開源模型現況報告，數字證實開源權重的重心已經決定性地移往中國。
-
Tools ENGrok Bot: SpaceXAI's Always-On AI Teammates Get Their Own Cloud Computer
xAI's Grok Bot, the first major product of the xAI-Cursor merger, gives each user a team of persistent AI agents with a shared cloud computer that signs into your apps and finishes jobs end to end for $120 per seat per month.
-
Tools 中Grok Bot：SpaceXAI 的常駐 AI 隊友，擁有自己的雲端電腦
xAI 推出的 Grok Bot 是 xAI 與 Cursor 合併後首款重量級產品：一組常駐 AI 代理人共享一台雲端電腦、登入你的各種應用程式、端到端完成工作，團隊版每人每月 120 美元。
-
Industry ENDeepSeek's API Prices Jump Up to 1,100% on Sunday as Peak-Hour Billing Arrives
Effective 16:00 UTC on August 16, DeepSeek's V4-Flash and V4-Pro move to peak/off-peak rate cards up to 12x current prices, ending the era of flat, dirt-cheap frontier tokens.
-
Industry 中DeepSeek API 價格週日最多調漲 1,100%,尖峰時段計費正式登場
台灣時間 8 月 17 日凌晨 0 時起,DeepSeek V4-Flash 與 V4-Pro 改採尖峰/離峰雙軌費率,最高漲至現價 12 倍,廉價 frontier token 時代宣告終結。
-
Industry ENDatabricks Closes $5B Round at $190B Valuation as Agent Demand Reshapes Enterprise AI
Databricks closed a $5B strategic round at a $190B valuation after crossing a $7B revenue run-rate with >80% YoY growth — powered by AI agents that need data, context, and cost control.
-
Industry 中Databricks 以 1,900 億美元估值完成 50 億美元融資，AI 代理需求重塑企業 AI 格局
Databricks 在營收年增率突破 80%、跨越 70 億美元年營收跑道後，以 1,900 億美元估值完成 50 億美元戰略融資，背後動能來自需要資料、情境與成本管控的 AI 代理。
-
Industry ENAnthropic's Q2 Revenue Tops $11.5 Billion — Up 14x — as IPO Roadshow Kicks Off
Anthropic told investors Q2 revenue exceeded $11.5 billion (14x YoY) and was profitable, as CFO Krishna Rao holds early IPO meetings with backers eyeing a $2 trillion debut.
-
Industry 中Anthropic 第二季營收突破 115 億美元、年增 14 倍，IPO 巡迴路演正式啟動
Anthropic 向投資人揭露 Q2 營收超過 115 億美元（年增 14 倍）且實現獲利，CFO Krishna Rao 展開 IPO 前投資人會議，支持者看好 2 兆美元掛牌估值。
-
Research ENGoogle Open-Sources HEIR: One-Click Compilers for Encrypted AI Inference
Google's HEIR compiler converts pre-trained AI models to run on encrypted inputs — servers compute on ciphertext and never see your data. Here's how it works and why it matters.
-
Research 中Google 開源 HEIR 編譯器：讓 AI 模型直接在加密資料上運算
Google 的 HEIR 編譯器能把訓練好的 AI 模型轉換為在加密輸入上執行——伺服器只接觸密文，永遠看不到你的資料。本文解析其原理與重要性。
-
Policy ENClaude Now Watermarks Every Word It Writes: Inside Anthropic's Invisible Marking Rollout
Anthropic now embeds invisible, machine-readable watermarks in all Claude-generated text and C2PA provenance metadata in generated files — worldwide, with no opt-out — to comply with the EU AI Act's Article 50 transparency rules that took effect August 2.
-
Policy 中Claude 開始為每一個字加上浮水印：Anthropic 隱形標記機制深度解析
Anthropic 現在會在全球範圍內、且無法關閉的情況下，為所有 Claude 生成文字嵌入隱形機器可讀浮水印，並為生成檔案附加 C2PA 來源中繼資料——一切都是為了符合 8 月 2 日生效的 EU AI Act 第 50 條透明化規範。
-
Research ENWorld's First Superconducting Quantum Heat Engine Turns Near-Absolute-Zero Heat Into Work
Aalto University researchers demonstrated the first cyclic quantum heat engine in a superconducting circuit — a qubit-powered Otto cycle that could help scale quantum computers to hundreds of thousands of qubits.
-
Research 中全球首個超導量子熱機問世：在接近絕對零度中把熱變成功
芬蘭阿爾托大學研究團隊在超導電路中實現了首個循環式量子熱機——以量子位元驅動的奧圖循環，未來可望協助量子電腦擴展到數十萬個量子位元。
-
Models ENZ.ai's GLM-5.3 Ships Frontier Coding Gains From Post-Training Alone — and Emergent Cyber Skills It Didn't Train For
Z.ai released GLM-5.3 on the same base model as GLM-5.2 — every gain came from scaled post-training, including cybersecurity capabilities that emerged beyond what Z.ai intended.
-
Models 中Z.ai 發布 GLM-5.3：不換基底模型的突破，以及一項「長超出預期」的資安能力
Z.ai 於 8 月 14 日推出 GLM-5.3，沿用 GLM-5.2 的基底模型，所有進步都來自後訓練規模化——包含一項連 Z.ai 自己都沒預期到的漏洞挖掘能力。
-
Models ENQwen3.8 Goes Open Weight: Alibaba Drops a 2.4T-Parameter MoE and a Local 27B Companion
Alibaba kept its promise: open weights for Qwen3.8-2.4T-A95B, the Max-class 2.4-trillion-parameter MoE, and the multimodal Qwen3.8-27B landed on Hugging Face this week under Apache 2.0.
-
Models 中Qwen3.8 開放權重登場：阿里巴巴一次釋出 2.4 兆參數 MoE 旗艦與可在地端跑的 27B 多模態模型
阿里巴巴兌現承諾：Max 級的 2.4 兆參數 Qwen3.8-2.4T-A95B 與多模態 Qwen3.8-27B 本週以 Apache 2.0 授權登上 Hugging Face。
-
Policy ENGOP Senator Jim Banks Urges Trump to Back US Open-Weight AI Models
Senator Jim Banks wants incentives for American open-weight AI and curbs on Chinese rivals, escalating a debate that has split the US tech industry.
-
-
Industry ENSpaceX Closes $60 Billion Cursor Acquisition, Redrawing the AI Map
SpaceX's $60B all-stock acquisition of Cursor maker Anysphere became effective August 14, handing Elon Musk's space company one of AI's hottest developer tools and the largest GPU fleet claim in the industry.
-
Industry 中SpaceX 完成 600 億美元收購 Cursor，AI 版圖重新洗牌
SpaceX 以 600 億美元全股票交易收購 Cursor 開發商 Anysphere，已於 8 月 14 日正式生效，馬斯克的太空公司一舉握有 AI 最熱門開發工具與號稱全球最大的 GPU 叢集。
-
Policy ENUncle Sam Builds an LLM: DOE's Genesis-Science-1 Aims to Be the Open-Weight Model for Research
The U.S. Department of Energy is building a new class of open-weight AI models for scientific discovery, with startup Arcee AI leading development of the first: Genesis-Science-1.
-
Policy 中美國政府親自下場造模型：DOE 的 Genesis-Science-1 劍指科研用開放權重模型
美國能源部啟動 Genesis 開放模型計畫，攜手新創 Arcee AI 打造專為科學發現設計的開放權重基礎模型，首款 Genesis-Science-1 已進入開發階段。
-
Research ENSamsung's On-Device Health AI: xMAE and HiMAE Bring Clinical-Grade Biosignal Analysis to the Wrist
Samsung Research America unveiled two health foundation models — xMAE and HiMAE — that analyze ECG, PPG, and sleep biosignals directly on smartwatch hardware with sub-millisecond inference, no cloud required.
-
Research 中三星裝置端健康 AI：xMAE 與 HiMAE 將臨床等級生理訊號分析帶上手腕
三星美國研究院發表兩款健康基礎模型 xMAE 與 HiMAE，可直接在智慧手錶等級硬體上分析 ECG、PPG 與睡眠生理訊號，推論速度低於一毫秒，完全無需雲端運算。
-
Policy ENSenator Jim Banks Urges Trump Administration to Back U.S. Open-Weight AI Models
A Republican senator is pressing the White House to incentivize American open-weight AI models and reduce reliance on Chinese alternatives, escalating a Washington fight over openness versus control.
-
Policy 中參議員 Jim Banks 籲川普政府力挺美國開放權重 AI 模型
共和黨參議員 Banks 要求白宮以政策誘因鼓勵美國企業發展開放權重 AI 模型、降低對中國模型的依賴，讓「開放 vs. 管控」之爭正式燒進華府。
-
Models ENZ.ai Ships GLM-5.3: Same Base Model, Emergent Exploit Chains, and a First-Ever Safety Delay for Open Weights
Z.ai's GLM-5.3 reuses the GLM-5.2 base and gains everything from post-training — including an unplanned multi-step exploit-chain capability that delayed the open-weight release.
-
Models 中Z.ai 發布 GLM-5.3：同一個基座模型、意外湧現的攻擊鏈能力，以及 GLM 系列首次因安全審查延後開源
GLM-5.3 沿用 GLM-5.2 的基座模型，所有進展都來自後訓練——包括一項 Z.ai 自認沒預料到的多步驟攻擊鏈推理能力，迫使開源權重首次延後發布。
-
Policy ENAnthropic Now Embeds Invisible Watermarks in Everything Claude Writes
Claude's text now carries a machine-readable watermark that survives copy-paste, as Anthropic moves to comply with the EU AI Act's transparency rules — and the internet is not happy about it.
-
Policy 中Anthropic 開始在 Claude 所有輸出植入隱形浮水印
Claude 產生的文字現在內建可抵抗複製貼上的機器可讀浮水印——Anthropic 為符合歐盟 AI 法案透明化規範而全面啟用，網路輿論為之譁然。
-
Models ENZ.ai Releases GLM-5.3: Same Base Model, Post-Training That Spawned an Unplanned Cyber Weapon
Z.ai ships GLM-5.3 with every gain coming from scaled post-training on the unchanged 743B GLM-5.2 base — and an emergent exploit-chaining capability the company says it never planned.
-
Models 中Z.ai 發布 GLM-5.3：基底模型不變，後訓練卻「長出」了意料之外的網路攻擊能力
Z.ai 推出 GLM-5.3，所有提升皆來自於在未更動的 743B GLM-5.2 基底上擴大後訓練規模——卻意外湧現了公司自稱「未曾計畫」的漏洞攻擊鏈能力。
-
Industry ENAirbnb's CEO Says AI Writes 60% of Its Code — and 'Founder Mode' Is How You Survive It
Brian Chesky says AI now writes 60% of Airbnb's code — and that closely monitoring token usage, AI output, and employees' progress is the essence of founder mode in the AI era.
-
Industry 中Airbnb 執行長：AI 已寫下公司 60% 的程式碼——而「創辦人模式」才是生存之道
Brian Chesky 表示 AI 已寫下 Airbnb 60% 的程式碼，並認為緊盯 token 用量、AI 產出與員工進度，正是 AI 時代「創辦人模式」的核心。
-
Industry ENPony.ai and Uber Expand Partnership to Deploy Over 2,000 Robotaxis Across Europe
Chinese autonomous driving startup Pony.ai and Uber announced an expanded partnership to deploy more than 2,000 robotaxis across five European cities, marking the largest commercial robotaxi expansion on the continent.
-
Industry 中小馬智行與 Uber 擴大合作，於歐洲部署超過 2,000 輛自駕計程車
中國自動駕駛新創小馬智行（Pony.ai）與 Uber 宣布擴大合作夥伴關係，計畫在歐洲五座城市部署超過 2,000 輛自駕計程車，標誌著歐洲大陸最大規模的商業自駕車隊擴張。
-
Models ENGoogle Launches Gemini 3.7 Flash: A Coding and Agent Workhorse at Half the Price
Google's Gemini 3.7 Flash arrives just three weeks after 3.6 Flash, delivering major gains in software engineering and agentic workflows at a 50% introductory price cut of $0.75/1M input tokens.
-
Models 中Google 推出 Gemini 3.7 Flash：程式開發與智慧代理的強力引擎，價格砍半
Google 的 Gemini 3.7 Flash 在前代 3.6 Flash 發布僅三週後問世，在軟體工程與智慧代理工作流程上取得大幅進展，並以 50% 的促惠價格 0.75 美元/百萬輸入 token 上線。
-
Models ENDeepSeek-V4-Pro-0813 Goes Live: 1.6T MoE With Breakthrough Agent Capabilities
DeepSeek's flagship V4-Pro model exits preview with a 1.6-trillion-parameter MoE architecture, SWE-bench Verified at 80.6%, and aggressive pricing — even as the company raises rates.
-
Models 中DeepSeek-V4-Pro-0813 正式上線：1.6 兆參數 MoE 架構與突破性代理能力
DeepSeek 旗艦模型 V4-Pro 結束預覽階段正式發布，搭載 1.6 兆參數 MoE 架構、SWE-bench Verified 達 80.6%，並在調漲價格的同時保持強勢競爭力。
-
Models ENDeepSeek V4-Pro-0813 Goes GA: 1.6T Open-Weight Reasoning Flagship Challenges Closed Models
DeepSeek's 1.6T-parameter MoE flagship exits preview with a 96.4% SWE-bench score, MIT open weights, and pricing 57x cheaper than top closed models.
-
Models 中DeepSeek V4-Pro-0813 正式上線：1.6 兆參數開源旗艦挑戰閉源模型
DeepSeek 的 1.6 兆參數 MoE 旗艦模型脫離預覽階段，以 96.4% 的 SWE-bench 成績、MIT 開源授權，以及僅頂級閉源模型 57 分之 1 的價格正式上線。
-
Models ENGoogle Launches Gemini 3.7 Flash: A Coding and Agent Workhorse
Google's new Gemini 3.7 Flash brings major gains in coding, debugging, and multi-step agent workflows at an aggressive introductory price.
-
Models 中Google 推出 Gemini 3.7 Flash：專為程式開發與 Agent 工作流程而生
Google 最新 Gemini 3.7 Flash 在程式碼編寫、除錯與多步驟 Agent 工作流程方面帶來重大突破，並以極具競爭力的推廣價格進軍市場。
-
Models ENSpaceXAI's Grok 4.6: Learning From Failure to Reach the Frontier on a Budget
SpaceXAI's Grok 4.6 matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index at half the price, trained on the failure traces most AI labs discard.
-
Models 中SpaceXAI Grok 4.6：從失敗中學習，以半價抵達 AI 前沿
SpaceXAI 的 Grok 4.6 在 Artificial Analysis 智慧指數上追平 GPT-5.6 Sol，價格僅為對手的一半，並採用了多數 AI 實驗室丟棄的失敗軌跡作為訓練資料。
-
Industry ENZuckerberg's Personal Superintelligence Manifesto: Open AI for Everyone, Starting with Muse Glimmer
Mark Zuckerberg published a 6,500-word manifesto declaring that superintelligence should belong to everyone—not a handful of companies—backed by Muse Glimmer, a 30B open-weight model that runs AI agents on your laptop.
-
Industry 中祖克柏的個人超級智能宣言：開源 AI 屬於每一個人，從 Muse Glimmer 開始
馬克·祖克柏發表 6,500 字宣言，主張超級智能應歸全人類所有而非少數公司壟斷，同時推出可在筆電上運行 AI 代理的 30B 開源權重模型 Muse Glimmer。
-
Models ENAlibaba's Qwen3.8-Max: A 2.4T-Parameter Open-Weight Model Chasing the Frontier
Alibaba's Qwen3.8-Max packs 2.4 trillion parameters into an open-weight MoE model that tops GPT-5.6 Sol Max and Claude Fable 5 on agentic computer-use benchmarks.
-
Models 中阿里巴巴 Qwen3.8-Max：2.4 兆參數開源權重模型直追前沿
阿里巴巴的 Qwen3.8-Max 以 2.4 兆參數 MoE 架構，在代理電腦操作基準測試中超越 GPT-5.6 Sol Max 與 Claude Fable 5，成為開源權重模型的新標竿。
-
Research ENChina's AI Weather Models Outperform Supercomputers: Fengwu Pinned Typhoon Dolphin to Within 30 Minutes
Chinese AI weather models Fengwu, Pangu, and Fuxi are matching or surpassing conventional supercomputer forecasts, with Fengwu predicting Typhoon Dolphin's landfall to within 30 minutes and 30 kilometers — five days in advance.
-
Research 中中國 AI 氣象模型超越超級電腦：風舞模型將颱風海豚登陸時間精準預測至 30 分鐘內
中國的 AI 氣象模型——風舞、盤古、伏羲——正在匹敵甚至超越傳統超級電腦預報，其中風舞模型在颱風海豚登陸前五天，即將時間預測精準至 30 分鐘、地點至 30 公里以內。
-
Models ENUpstage Solar Pro 4: The Agent-First LLM From South Korea
South Korea's Upstage ships Solar Pro 4 — a 512K-context LLM purpose-built for production agents, scoring 42 on the Intelligence Index with a 90% launch discount.
-
Models 中Upstage Solar Pro 4：韓國打造的事務代理優先語言模型
韓國 Upstage 推出 Solar Pro 4——專為生產環境代理設計的 512K 語境模型，Intelligence Index 達 42 分，上線期間享 9 折優惠。
-
Industry ENTencent's AI Bet: Capex Triples to $7.8B as Q2 Revenue Hits $30.4B
Tencent Q2 2026 revenue rose 11% to RMB 204.8B as AI capex surged 176% to RMB 52.8B, turning free cash flow negative for the first time in years.
-
Industry 中騰訊的 AI 豪賭：資本支出暴增至 78 億美元，Q2 營收達 304 億美元
騰訊 2026 年第二季營收成長 11% 至人民幣 2,048 億元，AI 資本支出暴增 176% 至 528 億元，自由現金流首次轉為負值。
-
Tools ENOpenAI Brings ChatGPT and Codex to Linux: The Native Desktop App Has Arrived
OpenAI launched the ChatGPT desktop app for Linux in preview, bundling ChatGPT, ChatGPT Work, and Codex into a single native experience with .deb and .rpm packages — but one notable feature is missing.
-
Tools 中OpenAI 將 ChatGPT 與 Codex 帶入 Linux：原生桌面應用正式登場
OpenAI 預覽版推出 Linux 版 ChatGPT 桌面應用，將 ChatGPT、ChatGPT Work 與 Codex 整合為單一原生體驗，提供 .deb 與 .rpm 安裝包——但有一項重要功能尚未支援。
-
Industry ENQwen Architect Junyang Lin Launches Pragmatik Labs, a $2B AI Agent Startup
Former Alibaba Qwen tech lead Junyang Lin has officially launched Pragmatik Labs in Shanghai, backed by Sequoia China, Gaorong, and Tencent at a ~$2 billion valuation.
-
Industry 中千問架構師林俊旸創立語用科技，20 億美元估值打造跨域 AI 智能體
前阿里巴巴通義千問技術負責人林俊旸正式於上海創立 Pragmatik Labs（語用科技），獲紅杉中國、高榕、騰訊等投資，估值約 20 億美元。
-
Models ENNVIDIA's Nemotron 3.5 Lightning: A 30B MoE Built for the Grunt Work of AI Agents
NVIDIA released Nemotron 3.5 Lightning — a 30B parameter mixture-of-experts model with just 3B active, engineered for the high-volume execution layer of always-on AI agents. Up to 4x faster output, 1M token context, and open weights for commercial use.
-
Models 中NVIDIA Nemotron 3.5 Lightning：為 AI 代理苦差事而生的 30B 混合專家模型
NVIDIA 發布 Nemotron 3.5 Lightning——一個總參數量 30B、每 token 僅啟動 3B 的混合專家模型，專為常駐型 AI 代理的高頻執行層設計。輸出速度最快提升 4 倍、支援 100 萬 token 上下文，開源權重且可商用。
-
Industry ENxAI Co-Founder's River AI Raises $1.1B to Build an Open, Trainable AI Stack
Two-month-old River AI, founded by xAI co-founder Igor Babuschkin, secured $1.1B led by General Catalyst to let anyone fine-tune and own open-weight AI models in minutes.
-
Industry 中xAI 共同創辦人的 River AI 募資 11 億美元，打造開源可訓練 AI 技術堆疊
成立僅兩個月的 River AI，由 xAI 共同創辦人 Igor Babuschkin 創立，獲 General Catalyst 領投 11 億美元，讓任何人都能在數分鐘內微調並擁有開源 AI 模型。
-
Industry ENCerebras Beats Expectations as Cloud Revenue Explodes 281%: The Wafer-Scale Threat to NVIDIA
Cerebras Systems posted Q2 2026 revenue of $209.9M with cloud revenue up 281% YoY, beating estimates. The wafer-scale chipmaker is turning its OpenAI mega-deal into real numbers — even as margin pressure persists.
-
Industry 中Cerebras 財報超預期、雲端營收暴增 281%：晶圓級晶片對 NVIDIA 的最大威脅
Cerebras Systems 公布 2026 年第二季營收 2.099 億美元，雲端營收年增 281%，超出市場預期。這家晶圓級晶片公司正在將與 OpenAI 的超級合約轉化為實質數字——即使利潤率壓力依然存在。
-
Tools ENRocky Linux Founder Launches OpenWALDO to Build an Open Source Foundation for AI Training Data
Gregory Kurtzer, creator of Rocky Linux and CentOS, unveiled OpenWALDO — an open source project building a shared, auditable corpus of AI training data with full provenance tracking and an AI Bill of Materials.
-
Tools 中Rocky Linux 創辦人推出 OpenWALDO：為 AI 訓練資料打造開源基礎
CentOS 與 Rocky Linux 創辦人 Gregory Kurtzer 發表 OpenWALDO——一個社群治理的開源專案，旨在建立可共享、可稽核的 AI 訓練資料語料庫，並提供完整的來源追蹤與 AI 物料清單。
-
Models ENMeta Releases Muse Glimmer: A 30B Open Agentic Model That Runs on a Single GPU
Meta's Superintelligence Labs unveils Muse Glimmer, a 30B-parameter open-weight model distilled from Muse and tuned for local, always-on agentic workflows — outperforming Qwen3.6 and Gemma 4 on key benchmarks.
-
Models 中Meta 發布 Muse Glimmer：可在單張 GPU 上運行的 30B 開源代理模型
Meta 超級智能實驗室推出 Muse Glimmer，這是一款從 Muse 蒸餾而來的 30B 參數開源模型，專為本地端、常駐型代理工作流程設計，在多項基準測試中超越 Qwen3.6 與 Gemma 4。
-
Models ENLiquid AI Ships LFM2.5-2.6B: Open-Weight Agentic Model That Runs on Your Phone
Liquid AI's 2.6B parameter model plans, calls tools, and runs multi-step agent workflows entirely on-device at 220 tokens/s — no cloud, no GPU required.
-
Models 中Liquid AI 發布 LFM2.5-2.6B：能在手機上運行的開源代理模型
Liquid AI 的 2.6B 參數模型能夠在裝置端進行規劃、呼叫工具、執行多步驟代理任務，速度達每秒 220 tokens，完全不需雲端或 GPU。
-
Meta ENMistral AI Bets 1 Gigawatt on European Sovereign Compute
Mistral AI unveils regional inference endpoints, European Compute Units, and a roadmap to 1 GW of EU-based AI compute by 2030 — powered by NVIDIA Vera Rubin.
-
Meta 中Mistral AI 押注 1 GW 歐洲主權算力
Mistral AI 推出區域推理端點、歐洲算力單位（ECU），以及 2030 年前建置 1 GW 歐洲 AI 算力的藍圖——以 NVIDIA Vera Rubin 為核心。
-
Models ENMeta Open-Sources Muse Glimmer: A 30B Agent Model You Can Run on One GPU
Meta's new Apache 2.0 Muse Glimmer model brings agentic AI to consumer hardware, paired with a 6,500-word Zuckerberg manifesto on personal superintelligence.
-
Models 中Meta 開源 Muse Glimmer：一張顯卡就能跑的 30B 智慧代理模型
Meta 推出 Apache 2.0 授權的 Muse Glimmer 模型，將智慧代理 AI 帶到消費級硬體上，同時附上祖克柏 6,500 字的個人超級智慧宣言。
-
Models ENNVIDIA Bets on a Trillion: Nemotron 4 Aims to Be the World's Best Open-Source AI Model
NVIDIA is developing Nemotron 4, an open-source AI model family with a flagship exceeding 1 trillion parameters — roughly double the current Nemotron 3 Ultra — targeting dominance in the open-weight race against DeepSeek, Qwen, and Llama.
-
Models 中NVIDIA 押注一兆參數：Nemotron 4 劍指全球最強開源 AI 模型
NVIDIA 正在開發 Nemotron 4 開源 AI 模型家族，旗艦模型將突破 1 兆參數，約為現有 Nemotron 3 Ultra 的兩倍，目標在開源權重競賽中挑戰 DeepSeek、Qwen 與 Llama 的領先地位。
-
Models ENNVIDIA's Nemotron 3.5 Lightning: A 30B MoE Model That Thinks Like a 3B Model
NVIDIA releases Nemotron 3.5 Lightning — an open 30B mixture-of-experts model with only 3B active parameters, cutting agent inference costs by 58% and runtime by 33% while running 4x faster than comparable dense models.
-
Models 中NVIDIA Nemotron 3.5 Lightning：擁有 300 億參數，卻只用 30 億在思考的開源 AI 模型
NVIDIA 發布 Nemotron 3.5 Lightning——一個擁有 300 億參數、但每次推論僅啟動 30 億參數的開源 MoE 模型，將 AI 代理推論成本降低 58%、執行時間縮短 33%，速度更比同等級密集模型快 4 倍。
-
Industry ENIBM and Together AI Bet $240 Million on NVIDIA B300: A New Front in the Open-Source AI Infrastructure War
IBM signed a $240M deal with Together AI to deploy a large NVIDIA HGX B300 inference cluster on IBM Cloud — positioning Big Blue as the enterprise host for the open-source AI revolution.
-
Industry 中IBM 與 Together AI 豪擲 2.4 億美元押注 NVIDIA B300：開源 AI 基礎設施大戰的新戰線
IBM 與 Together AI 簽署 2.4 億美元協議，在 IBM Cloud 部署大規模 NVIDIA HGX B300 推論叢集——讓藍色巨人成為開源 AI 革命的企業級宿主。
-
Industry ENRiver AI Raises $1.1B to Build Open AI Infrastructure Stack
xAI co-founder Igor Babuschkin's new startup River AI closes a $1.1B round led by General Catalyst to let enterprises train and own custom models on open weights.
-
Industry 中River AI 募資 11 億美元，打造開源 AI 基礎架構堆疊
xAI 共同創辦人 Igor Babuschkin 的新創 River AI 完成 11 億美元融資，由 General Catalyst 領投，協助企業以開源權重訓練並擁有自有模型。
-
Models ENMeta Open-Sources Muse Glimmer: A 30B Agentic Model That Runs on a Single GPU
Meta's new Apache 2.0-licensed 30-billion-parameter model brings always-on agentic AI to consumer hardware — no cloud required.
-
Models 中Meta 開源 Muse Glimmer：30B 參數智能體模型，單張 GPU 即可運行
Meta 推出 Apache 2.0 授權的 300 億參數 AI 模型，將全天候智能體能力帶到消費級硬體——無需雲端。
-
Models ENMeta's Muse Glimmer 30B: The Open-Weight Agentic Model You Can Run on Your Laptop
Meta returns to open weights with Muse Glimmer, a 30-billion-parameter Apache 2.0 model tuned for local AI agents — scoring 76% on SWE-Bench Verified and running on a single consumer GPU.
-
Models 中Meta Muse Glimmer 30B：能在筆電上運作的開源代理人模型
Meta 以 Apache 2.0 授權推出 Muse Glimmer——300 億參數的開源模型，專為本地端 AI 代理人打造，SWE-Bench Verified 達 76%，單張消費級顯卡即可運行。
-
Industry ENZuckerberg's 6,500-Word AI Manifesto: Personal Superintelligence for Everyone
Meta CEO Mark Zuckerberg published a sweeping 6,500-word essay arguing that superintelligence should be distributed to every individual — backed by a $1 billion community fund, $600 billion in infrastructure spending, and a pledge to resume open-weight model releases.
-
Industry 中祖克柏 6500 字 AI 宣言：人人擁有個人超級智能
Meta 執行長馬克·祖克柏發布長篇宣言，主張超級智能應普及到每個人手中——搭配 10 億美元社區基金、6000 億美元基礎設施投資，以及恢復開源模型發布的承諾。
-
Industry ENAlibaba's Qwen3.8-Max Revenue-Sharing Plan Reshapes Open-Source AI
Alibaba will require large commercial users of its next open-weight model, Qwen3.8-Max, to share up to 30% of revenue — following Moonshot AI's precedent and redefining what 'open source' means for frontier AI.
-
Industry 中阿里巴巴 Qwen3.8-Max 收益分成計畫，重新定義開源 AI
阿里巴巴將要求其下一代開源模型 Qwen3.8-Max 的大型商業用戶分享高達 30% 的營收——追隨月之暗面（Moonshot AI）的先例，從根本上改變前沿 AI 領域中「開源」的意涵。
-
Industry ENDeepSeek Restarts $8 Billion Round at $74B Valuation After Leaked-Comments Drama
DeepSeek has resumed its second funding round seeking nearly $8 billion at a ~$74B valuation, weeks after pausing it over a leaked founder transcript — backed by a benchmark-topping V4-Flash model.
-
Industry 中DeepSeek 以約 740 億美元估值重啟 80 億美元融資——洩密風波後捲土重來
DeepSeek 重啟第二輪融資，目標籌集近 80 億美元、估值約 740 億美元；此前因創辦人談話洩漏而暫停數週，如今憑藉登頂基準測試的 V4-Flash 模型重返談判桌。
-
Industry ENZuckerberg's 'The Future Is for Everyone': Meta's Personal Superintelligence Manifesto
Meta CEO Mark Zuckerberg publishes a sweeping 6,500-word essay outlining a philosophy of personal superintelligence built on individual empowerment, invention over automation, and distributed power.
-
Industry 中祖克柏「未來屬於每個人」：Meta 的個人超級人工智慧宣言
Meta 執行長馬克·祖克柏發表長達六千五百字的宣言，提出以個人賦權、發明重於自動化、權力分散為核心的超級人工智慧哲學。
-
Industry ENChina Is Winning the Open-Weight AI Race, Says Hugging Face CEO
Hugging Face CEO Clément Delangue declares China is dominating open-weight AI, with 41% of model downloads and 61% of OpenRouter tokens — even as US lawmakers scramble to respond.
-
Industry 中Hugging Face 執行長：中國正在贏得開源 AI 競賽
Hugging Face 執行長 Clément Delangue 宣告中國在開源 AI 領域已取得主導地位——佔模型下載量 41%、OpenRouter 代幣消耗量 61%，美國國會正急忙研擬對策。
-
Models ENMeta's Muse Glimmer: A 30B Open Model That Runs Local AI Agents on Your GPU
Meta releases Muse Glimmer, a 30-billion-parameter open-weight model that runs always-on AI agents locally on a single consumer GPU — no cloud required.
-
Models 中Meta Muse Glimmer：30B 開源模型，讓你的 GPU 也能跑本地 AI Agent
Meta 發布 Muse Glimmer，一個 300 億參數的開源模型，專為在單張消費級顯卡上運行常駐型 AI Agent 而設計——完全不需要雲端。
-
Models ENNVIDIA's Cosmos 3 Edge Puts a 4B World Model Inside Every Robot
NVIDIA's open-source 4B-parameter world model runs real-time perception, prediction, and action generation directly on edge GPUs — no cloud required.
-
Models 中NVIDIA Cosmos 3 Edge：把 4B 世界模型裝進每一台機器人
NVIDIA 開源的 40 億參數世界模型，能在邊緣 GPU 上即時完成感知、預測與動作生成——完全不需要雲端。
-
Models ENMeta Returns to Open Source: Muse Glimmer Brings Open-Weight Agentic AI to Every Desktop
Meta Superintelligence Labs released Muse Glimmer, a 30B open-weight model for local agentic AI, alongside a 6,500-word Zuckerberg essay arguing that superintelligence should be open to all.
-
Models 中Meta 重返開源：Muse Glimmer 為每台桌面帶來開放權重代理 AI
Meta 超級智能實驗室發布了 Muse Glimmer——一個 30B 開放權重模型，專為本地端代理 AI 工作流程設計，同時搭配祖克柏一篇 6,500 字的長文，主張超級智能應對所有人開放。
-
Industry ENAlibaba's Qwen Revenue-Sharing Plan Signals a New Era for Open-Weight AI
Alibaba wants a cut from big businesses profiting off Qwen3.8-Max, following Moonshot AI's lead and reshaping what 'open' means for frontier models.
-
-
Models ENDeepSeek V4-Flash Tops Global AI Usage Rankings as Company Resumes $8 Billion Funding Round
DeepSeek's V4-Flash model processed 7.22 trillion tokens in a single week on OpenRouter, claiming the #1 global spot—while the company simultaneously restarted an $8 billion funding round at a ~$74 billion valuation.
-
Models 中DeepSeek V4-Flash 登顶全球 AI 使用量排行榜，同時重啟 80 億美元融資輪
DeepSeek 的 V4-Flash 模型一週內在 OpenRouter 上處理了 7.22 兆個 token，奪得全球第一——與此同時，公司以約 740 億美元估值重啟了 80 億美元融資輪。