The Quiet Engineer Who Now Owns Qwen: Alibaba Names Dayiheng Liu Head of Its LLM Crown Jewels

Six months after the abrupt departure of Qwen’s founding technical lead, Alibaba has finally given its most strategically important AI project a permanent boss. The company has appointed Dayiheng Liu, a senior AI researcher and longtime core member of the Qwen team, as head of the Qwen large language model project, according to two employees with knowledge of the matter, in a report first published by The Information and picked up by Reuters and Techmeme on September 21, 2026.

The timing is hard to miss. Alibaba’s annual Apsara Conference opens Tuesday, September 22 in Hangzhou, and on the conference’s official speaker list, Liu’s portrait now sits prominently right after Group Chairman Joe Tsai and CEO Eddie Wu — with his title listed plainly as “Head of the Qwen Large Language Model Project.” The appointment was effectively announced by seating chart.

Why this job matters more than it looks

Qwen is not just another model family in Alibaba’s portfolio. It is the open-weight champion of the Chinese AI ecosystem and one of the most downloaded model families in the world, spanning everything from compact 3B models that run on laptops to the 24T-parameter Qwen3.8-Max frontier releases that go toe-to-toe with GPT-class systems on benchmarks. Alibaba has bet a reported $53 billion cloud-plus-AI investment plan on the strategy, and Qwen’s open-weights, revenue-sharing ecosystem is the spearhead of that bet.

Which is why the leadership vacuum since March has been so costly in perception terms. Junyang Lin — Alibaba’s youngest principal engineer, who had driven Qwen’s rise since 2022 — resigned suddenly on March 3, 2026, a day after Qwen’s most successful model launch ever. His exit triggered a cascade of departures from senior researchers and forced CEO Eddie Wu to personally take charge of a newly formed foundation-model task force. Alibaba Cloud CTO Zhou Jingren assumed interim leadership of the team while Liu, the original pre-training lead, took over core R&D.

Employees described the situation to reporters at the time as a scramble: no clear owner, multiple rounds of reorganization, and pointed questions from Qwen staff to Alibaba management about resources and direction. Liu’s appointment, the employees said, “has brought more clarity to the Qwen team’s leadership after multiple rounds of reorganization earlier this year.”

Who is Dayiheng Liu?

Liu’s profile is a study in contrast with the celebrity researchers who front most frontier labs. He is not an overseas-returned PhD or a frequently-quoted public face. He earned his doctorate at Sichuan University’s DICALab, advised by Professor Jiancheng Lv, joined Alibaba in 2021 as an algorithm expert in DAMO, and has been a core R&D member of Qwen essentially ever since. His published work concentrates on the unglamorous engine room of LLMs: pre-training, reasoning, reinforcement learning, and natural language generation. His fingerprints are on generations of the family, including the Qwen2-VL vision-language models that helped establish Qwen’s international credibility among developers.

Notably, Liu was also part of Huawei’s “Genius Youth” talent development program earlier in his career — a detail that matters in China’s competitive AI labor market, where talent poaching between Huawei, Alibaba, ByteDance, and a wave of well-funded startups has become one of the defining storylines of the race.

He also showed his spine in March. During the tense all-hands-style meeting that followed Lin’s resignation, it was Liu — then the team’s reinforcement learning lead — who represented Qwen members in raising a series of pointed questions to Alibaba’s senior management. Now the person who asked management the hard questions is the person management has handed the team to.

What he inherits

The in-tray is not short. Qwen3.8-Omni-Flash and Qwen-Image-2.1 have kept the family in the news, but the competitive ground is shifting fast. DeepSeek’s founder has publicly committed to training on Huawei’s Ascend chips, threatening Qwen’s position as the default open-weight choice in the domestic ecosystem. Google’s AX agent infrastructure and a flood of agentic frameworks are changing what developers expect from a foundation model. And Alibaba’s own commercial machinery — the Token Foundry, the agent-native cloud suite announced at WAIC in July — needs Qwen releases to keep landing on schedule.

The Apsara Conference this week will be Liu’s first public test. The agenda promises “full-stack AI breakthroughs: Qwen foundation models, agent-native infrastructure, and real customer wins,” across three keynotes and more than 140 sub-forums. Whatever is announced on that stage Tuesday is now officially his product.

The bottom line

Alibaba’s choice signals continuity over star power. Rather than hiring a big name from outside or elevating a spokesperson-executive, it promoted the engineer who has been in the pre-training trenches since the beginning — a message aimed as much at Qwen’s remaining researchers, and at the open-source community that builds on Qwen daily, as at competitors. After six months of interim arrangements, the message is simple: Qwen has an owner again, and he’s someone who knows exactly where the bodies are buried in the training pipeline.

Whether quiet competence can hold the line against DeepSeek’s momentum and the talent drain that started in March is the open question. But for a company that spent half a year without a clear answer to “who runs Qwen?”, having an answer at all — right before 50,000 square meters of Apsara expo floor opens in Hangzhou — is the point.