“me stepping down. bye my beloved qwen.” In the early hours of March 4, that brief sentence appeared on social media and instantly set the AI world ablaze. The author was Lin Junyang, technical lead of Alibaba’s Tongyi Qwen team and the company’s youngest P10-level engineer. His near-wordless exit from what is arguably Alibaba’s most strategically valued AI business sent shockwaves through the industry.
From a Highlight Moment to an Abrupt Curtain Call
The sequence of events unfolded with striking speed. On March 2, Alibaba open-sourced the Qwen 3.5 small-model series — spanning the 0.8B, 2B, 4B, and 9B parameter ranges — to immediate fanfare. Elon Musk personally reposted the announcement, remarking on its “astonishing intelligence density.” The same day, Alibaba announced it would unify its enterprise large-model brand and consumer application brand under a single “Qwen” identity.
On March 3, Jack Ma convened Alibaba and Ant Group’s core leadership at Hangzhou Yungu School to discuss AI strategy. That same day, Lin Junyang formally submitted his resignation.
By the early hours of March 4, he had announced his departure publicly on X. That afternoon, he posted on WeChat Moments that he “really needed rest,” while separately reassuring colleagues: “Qwen brothers, keep doing it according to the original plan — no problem.”
The episode did not stand alone. Around the same period, Yu Bowen, head of Qwen post-training, officially left the company. Hui Bin, head of Qwen Code, had already departed in January to join Meta. Core contributors including Kaixin Li were also reported to have exited during the same window. Taken together, the departures amounted to what observers have called an unusual personnel earthquake.
A Collision Between Technical Ideals and Organizational Reality
Multiple sources point to the same trigger: organizational restructuring. According to reports, Tongyi Lab had been planning to dissolve the Qwen team’s vertically integrated structure and reorganize it into horizontally divided units covering pre-training, post-training, text, and multimodal work separately. Under that arrangement, Lin Junyang’s managerial authority would have been sharply curtailed.
Yet that may only be the surface explanation. A deeper fault line appears to run through the technical roadmap itself. Lin Junyang had long and publicly maintained that pre-training and post-training must be deeply coupled, and that tighter integration — extending to infrastructure and training teams — was essential for meaningful progress. Alibaba’s restructuring, by contrast, pointed toward breaking the team into a pipeline model, substituting organizational predictability for what critics might call individual-driven development.
A shift in performance metrics may have deepened the divide. Those familiar with the situation suggest Lin Junyang’s departure coincided with an internal transition away from technical benchmarks toward daily active user counts as the primary measure of success — a reorientation that a technical idealist would have found difficult to accept.
Hidden Pressures Behind the Open-Source Crown
Before assessing the fallout from this leadership shake-up, it is worth taking stock of where Qwen actually stands.
Under Lin Junyang, the achievements were hard to dispute. Global downloads exceeded one billion, surpassing Meta’s Llama to rank as the world’s leading open-source model family. Derivative models built on Qwen’s foundation topped 200,000, accounting for more than 40 percent of new language models on the Hugging Face platform. At the 2025 GTC conference, NVIDIA CEO Jensen Huang presented data indicating that Qwen already commanded the majority of the open-source model market. A Sullivan report found that in the second half of 2025, average daily enterprise-level large-model invocations in China reached 37 trillion tokens, with Qwen holding a 32.1 percent share — ranking first.
The closed-source flagship track tells a more complicated story. Qwen3-Max, released in 2025, exceeded one trillion parameters and outperformed mainstream international models on evaluations including GPQA. But the broader industry view is that Qwen has begun to show signs of strain in its pursuit of the top tier of closed-source models — GPT-5 and Gemini 3 among them.
Resource constraints compound the pressure. The Qwen team supporting updates across more than a hundred model variants numbers only around 100 people. Tongyi Lab in its entirety has only just exceeded 600. ByteDance’s Seed team, responsible for foundation-model training, operates at roughly 1,000.
The commercialization challenge may be the most structurally difficult of all. Alibaba has pursued a strategy of offering models free of charge while monetizing cloud services — an effort to position Qwen as the industry’s default infrastructure, a Linux for the server era. The costs of that approach, however, are considerable. Meta spent billions of dollars training Llama before releasing it at no charge, and that expenditure remains largely invisible in financial statements. As Alibaba has tilted its AI priorities toward the consumer market — specifically the Qwen App — a widening gap has opened between the foundation-model team’s objectives and the group’s commercial direction. The departure of Lin Junyang and others reflects, in no small part, the friction generated by that divergence.
Wolves on Every Side
Lin Junyang’s exit arrived at the most competitive moment yet in China’s large-model race.
On the domestic front, a once three-way contest has given way to a broader free-for-all. ByteDance’s Doubao has leveraged the Douyin ecosystem and aggressive AI investment to build a commanding position in the consumer market. DeepSeek ignited the competitive landscape with its R1 reasoning model, then deepened its advantage by adapting closely to domestic compute infrastructure. Tencent’s Yuanbao operates on a dual-model architecture combining Hunyuan T1 and DeepSeek; Tencent has not yet deployed its full resources, but its ownership of the WeChat ecosystem remains a formidable variable. Meanwhile, Zhipu and MiniMax — both now listed in Hong Kong — have been pressing forward with considerable momentum.
Internationally, open-source models are rewriting the competitive calculus. Jensen Huang stated plainly at the 2025 GTC conference that open-source models had become powerful enough to materially accelerate AI adoption. A joint report by MIT and Hugging Face found that, over the past year, China-developed models accounted for 17.1 percent of global open-source model downloads, narrowly ahead of U.S.-developed models at 15.8 percent. The success of Qwen and DeepSeek has demonstrated that open-source strategies can erode the dominance of closed-source incumbents. The risk, equally clear, is that personnel turbulence could slow Qwen’s iteration pace and erode the ecosystem advantage it has worked to build.
Where Does Qwen Go From Here?
On the afternoon of March 4, Alibaba Group CEO Eddie Wu held an emergency all-hands meeting, apologized to Qwen employees, and stated that the team was expanding rather than contracting. Market confidence, however, had plainly taken a hit.
The near-term consequences are difficult to avoid. Core departures risk disrupting technical continuity, damaging team morale, and introducing a period of operational uncertainty. The longer-term trajectory will depend on how effectively Alibaba resolves competing internal priorities — whether it can grant a new leadership team genuine technical decision-making authority, whether it can identify a sustainable business model, and how it proposes to retain key talent at a time when ByteDance, Zhipu, and others are offering aggressive equity packages.
Yu Bowen’s reported successor is Zhou Hao, a former senior staff researcher at DeepMind who joined Alibaba earlier this year. Hao Zhou, a former researcher on Google’s Gemini team, is said to be a possible addition as well. Whether this new generation of leadership will preserve Qwen’s open-source character — or steer it toward more explicitly commercial ends — remains an open question.
Lin Junyang’s departure marks more than a career transition. It signals, in the view of many observers, the close of the first chapter of China’s large-model era. The competition has entered its second half — a contest defined not by research breakthroughs alone, but by compute capacity, capital deployment, ecosystem control, and business model viability. When technical ideals run up against organizational realities, and when open-source principles collide with commercial KPIs, individual conviction rarely holds against institutional momentum.
Lin Junyang has not announced his next move. For a technologist of his caliber, options will not be scarce. For Alibaba Qwen, the harder test is only now beginning: how to sustain technological leadership without its animating figure, and how to find a workable equilibrium between open-source ambition and commercial necessity. The answers will go a long way toward determining whether Qwen remains a central player in the next phase of the AI era.
In a field that reinvents itself weekly, nothing stays fixed for long.
[Disclaimer]: The above content reflects analysis of publicly available information, expert insights, and BCC research. It does not constitute investment advice. BCC is not responsible for any losses resulting from reliance on the views expressed herein. Investors should exercise caution.