On March 4, 2026, Lin Junyang, the technical architect behind Alibaba’s open‑source Qwen models, posted a terse social‑media note announcing his resignation. His message unleashed a cascade of departures from the Qwen team over the following hours, including senior engineers responsible for post‑training and key contributions to Qwen3.5’s multimodal and coding capabilities. The sudden turnover has become the first major personnel shock to China’s AI sector after the Lunar New Year and has prompted urgent internal damage control at Alibaba Cloud’s Tongyi Lab.
Lin, born in 1993 and a Peking University computer science alumnus, rose rapidly at DAMO Academy after joining in 2019, later becoming the technical lead of Tongyi Qwen when multiple research teams were folded into Alibaba Cloud in 2022. Under his stewardship, the Qwen family of models — especially a set of compact open‑source variants — built a global developer following: more than 400 models, roughly 200,000 derivatives and over one billion downloads. He was promoted to one of Alibaba’s youngest P10s after Qwen’s breakout success in 2025, but his departure comes less than a year into that rank and less than a week after Qwen3.5’s high‑profile release.
Company leaders held an emergency All Hands on the afternoon of March 4 to stabilise staff and recast the reshuffle as expansion rather than contraction. Senior executives framed the reorganisation as a move to split Qwen’s development into horizontally specialised teams — pre‑training, post‑training, text, and multimodal — mirroring approaches used at other Chinese tech firms. That structural change brought new hires, including an ex‑DeepMind reinforcement‑learning lead, into roles that now report directly to Tongyi’s head rather than to Lin, shrinking his operational remit and unsettling the project’s prior model of vertical integration.
The proximate fault line is performance: Qwen’s small open models won plaudits in developer communities and on platforms such as Hugging Face, where the Qwen3.5 compact variants dominated trend charts and drew praise from Elon Musk. But the flagship Qwen3.5‑397B did not meet internal expectations. In benchmark rankings such as LMArena’s overall list, the 397B model sat well below the top tier, and specialist Chinese language leaderboards also placed it behind domestic rivals. Insiders say management weighed that mixed performance against the costs of maintaining a large open‑source footprint and concluded organisational change was necessary.
That calculus points to a deeper strategic tug‑of‑war inside Alibaba: should the group prioritise broad community impact through open‑source models that lower barriers for startups and researchers, or should it accelerate monetisation through cloud services, consumer products and tightly controlled model IP? Alibaba’s public AI strategy stitches together base models, cloud compute and in‑house chips, but executives have made clear that open‑source influence alone is an incomplete route to sustained revenue. The reorganisation appears designed to align model development more directly with product and cloud business lines.
For the wider ecosystem the stakes are concrete. Thousands of small and medium AI firms have leveraged Qwen’s compact releases because they are low‑cost, easy to deploy and well documented; many entrepreneurs now worry that any tilt away from open distribution would raise their operating costs or force them to migrate. Conversely, talent churn at the core of a flagship project can slow innovation even if the short‑term engineering teams remain intact, and hiring senior external leads signals that Alibaba is ready to import different engineering philosophies to chase performance or product fit.
In the near term, insiders and observers judge the operational damage to Alibaba’s AI roadmap as limited: Lin’s team achieved the crucial zero‑to‑one breakthroughs and a large open ecosystem remains in the wild. But the departure is a vivid symbol of the unresolved tension between community‑oriented, long‑horizon R&D and the quarterly discipline of a public company. How Alibaba reconciles those priorities will shape not only the Qwen programme’s technical trajectory but also the competitive landscape for open‑source AI in China.
The episode echoes a broader industry realignment: several global AI leaders are reassessing open‑source strategies as compute costs and commercial imperatives mount. Whether Alibaba doubles down on open ecosystems as a strategic differentiator, or moves incrementally toward tighter control and clearer monetisation, will determine whether Qwen remains the developer staple that nurtured scores of startups or becomes another advanced model engineered primarily for enterprise revenue.