Four Thousand Pixels and a Memory: Inside Tencent's Hy Image 3.5 Preview

Four Thousand Pixels and a Memory: Inside Tencent’s Hy Image 3.5 Preview

On September 22, 2026, while Alibaba CEO Eddie Wu was laying out a 5–10 trillion-parameter model roadmap and a new Zhenwu V900 accelerator on stage at the Apsara conference in Hangzhou, Tencent quietly shipped its own counter-move — not on a keynote stage, but in a cloud API listing. Hy Image 3.5 Preview, the latest generation of the company’s Hunyuan image model family, went live under the model ID hy-image-v3.5-preview in Tencent Cloud documentation updated September 21, and it is now rolling into the Yuanbao chatbot, Tencent’s film-editing suites, and its design tools.

The timing was deliberate. Bloomberg framed the release plainly: Tencent is shipping an AI image model to catch ByteDance and Alibaba, its two closest rivals in China’s escalating multimodal race. But the more interesting story is not the competitive posturing — it is what the model is actually designed to do, because Hy Image 3.5 Preview is built around a fundamentally different assumption about how people want to create images.

An image model that remembers

Most image generators treat every prompt as a fresh start. You describe something, the model produces a picture, and if you want a change you either re-roll the whole thing or wrestle with inpainting masks. Hy Image 3.5 Preview inverts that model: it is designed for revision loops, not one-shot generation.

The API exposes a chat-style messages format. A request can carry text plus a reference image; the model returns a result; and on the next turn, the developer appends the previously generated image and the conversation history to the request. The model carries that context forward, so a targeted instruction — “make the jacket red,” “move the subject left,” “soften the background lighting” — modifies the existing picture rather than producing an unrelated new one.

This sounds like a small interface detail. It is not. Preserving a prepared history across turns makes sequential editing part of the interface design rather than an undocumented behavior that application developers must reconstruct themselves. For teams building creative workflows — a design tool iterating on a mockup, a film-editing pipeline adjusting a storyboard frame, a marketing team refining a campaign visual — the difference between “regenerate and hope” and “revise with memory” is the difference between a toy and a tool.

The synchronous return path matters too. Rather than submitting a job and polling for completion, applications receive the completed image in the same request. For interactive editing loops where the user is watching the screen and waiting for the next revision, latency behavior of this kind is the product.

What the API exposes

According to the Tencent Cloud documentation, the service supports:

• Text-to-image and image-to-image generation, with reference images accepted as input for guided results.
• Multi-turn editing through the chat-style messages format, with prior generated images and conversation history appended to subsequent requests.
• Custom image sizing, including explicitly requested 4K output at up to 4,096 × 4,096 pixels.
• Optional external-search enhancement, letting the model pull in outside context when generating.

That last item is easy to overlook. Search-augmented image generation — where the model grounds what it draws in retrieved references rather than frozen training data — is becoming a quiet differentiator across the industry, and its presence in the API suggests Tencent expects developers to build factually anchored visual workflows, not just stylized ones.

The claims, and what they’re worth

Tencent says it tested the model with hundreds of in-house designers. In those internal trials, the company reports outputs comparable to ByteDance’s Seedream 5.0 Pro — currently one of the strongest image models shipping out of China — and slightly ahead of Google’s Nano Banana Pro and Alibaba’s Qwen-Image-3.0 Pro.

Those are aggressive claims, and it is worth being precise about their evidentiary status: Tencent published no quantified metrics alongside the release. No benchmark tables, no human-preference study sizes, no win-rate percentages. The competitive picture exists, for now, only as the company’s own assessment of its own model, evaluated by its own employees. The share price liked it anyway — Tencent stock rose more than 7% in Hong Kong after the announcement — but the market reacting to a claim is not the same as the claim being verified.

The real test will be twofold. First, whether independent evaluations — the image-model leaderboards and community blind tests that have become the de facto arbiters in this space — reproduce Tencent’s ranking once the preview is widely accessible. And second, whether the revision workflow holds up in production: multi-turn image editing is notoriously hard to keep stable, with identity drift and composition degradation accumulating across successive rounds. A model that edits beautifully on turn two and falls apart by turn five has not solved the problem the API is designed around.

The full-stack shadow war

The release also lands in a broader strategic context that Bloomberg’s framing only hints at. Alibaba used the same 24 hours to unveil a chip (Zhenwu V900, claiming triple the performance of its predecessor) and commit to 20 GW of data-center capacity by 2032. ByteDance’s Seedream line continues to set the pace on image quality among Chinese labs. Tencent’s answer — shipping an editing-first image model directly into consumer and professional products simultaneously, backed by a cloud API — reflects its traditional playbook: win through distribution.

Yuanbao gives Hy Image 3.5 an immediate consumer surface. The film-editing and design tool integrations give it professional ones. And the synchronous, history-aware API gives third-party developers a reason to build on Tencent Cloud rather than treat Hunyuan as a research curiosity. That is the same funnel logic — product integration driving cloud consumption — that has defined Tencent’s AI strategy since the Hunyuan reorganization.

Whether it works this time depends on the one thing press releases cannot manufacture: whether the model, stripped of its own marketing, actually edits as well as it generates. The preview tag in hy-image-v3.5-preview is honest in that respect. This is a claim awaiting verification, shipped in the most verifiable way possible — into products where millions of users will find out.

For now, the scoreboard reads: Tencent has made image editing a first-class citizen of its frontier image model, at resolutions up to 4K, with an API designed for exactly the iterative workflows that creative teams actually run. If the internal designer trials translate to external reality, ByteDance and Alibaba have a real problem. If they don’t, this becomes another chapter in the long history of labs benchmarking their own homework.