In the early hours of March 4, 2026 — Beijing time, so it was still the evening of March 3 for most of the Western AI community — Lin Junyang posted five words on X that hit the global open-source world like a small earthquake. “Me stepping down. bye my beloved qwen.” No justification. No long statement. Just that, from the 32-year-old technical lead who had steered Alibaba’s Qwen model series from a nascent internal project to one of the most downloaded open-weight AI model families on the planet, with over 600 million downloads.
The post received over 5,000 likes in a matter of hours. Hundreds more responses flooded in, expressing gratitude and, to be honest, grief. One team member posted: “Qwen is nothing without its people.” The phrase probably didn’t fall as smoothly as it should have.
It felt more than just a personnel change because Lin’s departure was not a unique incident. Hui Binyuan, who led the Qwen Code division, had quietly left Alibaba in January 2026 to join Meta — the same Meta that Qwen had been increasingly outpacing in the open-source leaderboards. Yu Bowen, head of post-training, announced his own resignation the same day as Lin.
Three architects of Qwen’s most celebrated capabilities — coding, post-training alignment, and overall technical direction — gone in less than two months. A research scientist named Wenting Zhao called Lin’s exit “the end of an era.” That description, from someone inside the project, carried a weight that Alibaba’s subsequent communications struggled to match.
What triggered the departures is easier to sketch than to confirm with precision, partly because none of the three men gave detailed explanations and partly because Alibaba, in the Chinese corporate tradition, has said very little publicly about internal dynamics. What is known is that Qwen was made a group-level strategic priority, which resulted in changes to the organizational structure and an increase in the foundation model team, which Lin was allegedly unable to accept.
This cycle seems almost inevitable: the researcher who created something from the ground up and is familiar with its design in a way that no one promoted to supervision will ever fully understand is given a reorganized position that requires him to manage rather than create. Every industry experiences it. It hurts especially in the field of artificial intelligence, where the gap between those who create the models and those who manage the company has historically been quite small and is growing.
Here, the larger background is important and worth considering. The Qwen team at Alibaba was working in a Chinese tech environment while under constant pressure from many sources. Chinese AI firms’ domestic price battles have drastically reduced profits, forcing businesses to pursue commercialization schedules that don’t necessarily coincide with the cycles of basic research.
Additionally, Beijing has reportedly started forcing prominent AI researchers at large private companies, such as Alibaba, to get official approval before traveling abroad. Beijing is increasingly regarding advanced AI as a subject of national security equivalent to defense technology. That is a significant limitation for top engineers who work internationally, go to international conferences, and pursue research wherever it takes them. It’s a ceiling, and ceilings tend to force people in the direction of exits over time.
Alibaba responded quickly and intelligently. A Foundation Model Task Force was established by CEO Eddie Wu, who personally oversaw the endeavor. Zhou Jingren, CTO of Alibaba Cloud, took over the Tongyi Lab. Zhou Hao, a former senior researcher at Google DeepMind, was hired to oversee post-training, the particular role that Yu Bowen had left empty.
The business revealed ambitions to hire top people and increase funding for AI development. These actions were taken promptly and are the appropriate ones to take in a situation like this. It’s still uncertain if organizational authority and speed can take the place of Lin Junyang and his colleagues, who were more than just participants to the project—they were the ones who understood why Qwen functioned as it did. A task force memo does not convey that type of institutional knowledge.
