Junyang Lin, the 32-year-old technical lead behind Alibaba's Qwen model family, announced on X on March 4 that he was stepping down. His post on X was brief and blunt: "me stepping down. bye my beloved qwen." Within hours, staff research scientist Binyuan Hui, contributor Kaixin Li, and post-training head Yu Bowen had all signaled their own exits. The departures came less than 24 hours after the team shipped Qwen 3.5 Small, a four-model series that Elon Musk praised on X for its "impressive intelligence density."
What actually happened
According to Chinese outlet LatePost's reporting, Lin submitted his resignation on March 3. The timing was not coincidental. Alibaba's Tongyi Lab had been planning to break up the Qwen team's vertically integrated structure, splitting it into horizontal units for pre-training, post-training, text, and multimodal work. Lin, who ran Qwen as a self-contained operation covering everything from infrastructure to multimodal research, reportedly opposed the fragmentation.
The restructuring also brought in Zhou Hao, a former senior staff research scientist at Google DeepMind who contributed to Gemini 3.0, AI Mode, and Deep Research. South China Morning Post reported that Zhou joined Alibaba earlier in 2026 as head of post-training research, replacing Yu Bowen and reporting directly to Alibaba Cloud CTO Zhou Jingren. Lin apparently did not accept the new arrangement easily.
There were scope conflicts too. Over the past year, Lin had expanded Qwen's work into image generation, voice models, and robotics, areas that overlapped with other teams inside Tongyi Lab. Corporate leadership wanted modular teams. Lin wanted tight integration. Someone had to lose.
The damage control meeting
An all-hands within Tongyi Lab followed almost immediately, and by multiple accounts it was tense. Alibaba CEO Eddie Wu apologized for poor communication around compute resource constraints but insisted Qwen remained his top priority. Cloud CTO Zhou Jingren admitted resources had been tight and, in a striking admission, suggested he too had been marginalized internally.
When a team member asked whether Lin could return, the Chief HR Officer shut it down. The company "cannot put him on a pedestal" and would not "accept irrational demands that spare no cost to retain him," according to VentureBeat's account of the meeting. That phrasing is revealing. It suggests the departure involved more than a quiet disagreement about org charts.
Why it stings
Lin was not just a team lead. He was Alibaba's youngest P10 executive, the company's senior leadership tier. He joined in 2019 out of Peking University, worked on the M6 and OFA models, and turned Qwen from an internal side project into something with over 700 million cumulative downloads on Hugging Face and more than 180,000 community-built derivative models, surpassing even Meta's Llama on that metric. His Google Scholar profile shows over 42,000 citations. The Qwen3 technical report alone has nearly 9,000.
Colleague Wenting Zhao called it "the end of an era." Contributor Chen Cheng was more direct, writing on X that he was "heartbroken" and adding: "I know leaving wasn't your choice." Xinyu Yang, a researcher at DeepSeek, posted a sharper critique, calling the move a replacement of an excellent leader with someone from Google Gemini, driven by DAU metrics. "If you judge foundation model teams like consumer apps, don't be surprised when the innovation curve flattens."
That DAU pressure is worth pausing on. Alibaba launched a consumer-facing Qwen app in November 2025 and has been consolidating AI efforts into a "Qwen C-end Business Group" that merges model labs with consumer hardware teams for AI glasses and rings. The app's monthly active users reportedly jumped from 31 million in January to 203 million in February, numbers that sound enormous but deserve scrutiny. MAU spikes around product launches are common and often don't stick.
So what comes next
Alibaba's stock fell as much as 5.3% in Hong Kong on the news, its largest intraday drop since October, though broader market conditions also contributed. The models themselves keep shipping. Qwen 3.5 Small is live on Hugging Face and ModelScope, and the GitHub repos remain active. Short-term continuity looks fine.
The medium-term is murkier. The simultaneous exit of a tech lead, a post-training head, a research scientist, and a contributor does not happen by accident. Plans for a Qwen international base in Singapore may now be shelved, according to details noted by Kaixin Li. And while Alibaba insists this is an expansion, not a contraction, the pattern is familiar. DAMO Academy lost 30% of its staff in 2022. Key Qwen names left in 2024 and early 2025. This is the third wave.
Zhou Hao brings legitimate credentials from DeepMind. But the open-source community trusted Lin specifically, and that trust does not transfer with a job title. Alibaba has not named a successor for the overall tech lead role. The next concrete signal will be whether Qwen's release cadence and open-weight commitment hold through Q2 2026.
