Lin Junyang, the technical lead of Alibaba's Qwen team, unexpectedly announced his resignation on social media early March 4. In a post, he wrote: "me stepping down. bye my beloved qwen." It was later confirmed that Lin had formally submitted his resignation to Alibaba on the afternoon of March 3. The news was subsequently shared within the Qwen team, with some members reportedly becoming emotional upon hearing it.
By 4 a.m. Beijing time, Lin’s social media post had garnered over 5,000 likes and more than 700 comments, many expressing gratitude for the Qwen team’s contributions to open-source large language models.
On the same day, Yu Bowen, the post-training lead for Qwen, also officially left the company. His responsibilities will be taken over by Zhou Hao, a former senior researcher at DeepMind who joined Alibaba’s Tongyi Lab earlier this year. Zhou Hao reports to Zhou Jingren, Alibaba Cloud’s CTO and head of Tongyi Lab.
Additionally, Hui Binyuan, the former lead of Qwen Code, left Alibaba in January 2026 to join Meta. Lin Junyang had taken over Qwen Code responsibilities after Hui’s departure, and as recently as last week, Lin was still sharing recruitment information related to Qwen Coding Agent on social media.
Multiple sources close to the matter described Lin’s departure as sudden, expressing sentiments such as "regretful" and "he truly loved Qwen." His decision is believed to be linked to an ongoing organizational restructuring within the Qwen team.
The Qwen team, which Lin directly managed, operates under Tongyi Lab, led by Zhou Jingren. Recently, Tongyi Lab has planned to split the Qwen team from a vertically integrated structure—covering various training processes and modalities—into horizontally divided teams focusing on areas like pre-training, post-training, text, and multimodal tasks. These teams will remain under Tongyi Lab, but Lin’s managerial scope was reduced.
This shift toward decentralization did not align with Lin’s view of technological trends. Since last year, he had repeatedly emphasized the need for closer integration and communication among pre-training, post-training, infrastructure, and training teams. It was previously reported that, starting mid-last year, the Qwen model team began building its own infrastructure unit, a function previously handled mainly by Alibaba Cloud’s AI platform PAI, which supports infrastructure needs across various teams within Tongyi Lab.
Over the past two years, several major Chinese tech companies have repeatedly adjusted their AI team structures. Alibaba had been relatively stable in this regard. For comparison, ByteDance’s Seed model team employs a "horse race" approach with multiple teams working on the same direction or modality, while its Doubao main model series is divided by pre-training and post-training processes. Tencent, after adjustments last year, adopted a more integrated structure, with both model training and infrastructure teams reporting to Yao Shunyu.
Given Qwen’s reputation and achievements in recent years, Lin Junyang is not short of opportunities. Several investors and large companies had previously approached him, with some encouraging him to start his own venture and others extending job offers.
Prior to these changes, the Qwen team faced subtle tensions within Alibaba. On one hand, Qwen has garnered a large following among global open-source developers. Its diverse model sizes are popular with small and mid-sized startups, and companies like Cursor fine-tune and further train based on the Qwen series. Qwen’s multimodal open-source models are also chosen as base models by many Chinese embodied AI companies.
Simultaneously, Qwen and Lin continuously expanded their capabilities, leading to overlaps with other parallel teams within Tongyi Lab. For instance, Qwen was also developing a vision-language-action embodied model, an area also covered by a team led by Xu Zhuhong within Tongyi Lab. Qwen was also working on text-to-image models and speech models, overlapping with Tongyi Lab’s Tongyi Wanxiang and Bailing teams. As Qwen began building its own infrastructure team, it increasingly resembled a comprehensive, full-stack AI lab.
On the other hand, Alibaba internally continued to evaluate Qwen’s output and value. This included questioning the commercial efficiency of open-source models: while Qwen is highly regarded, its open-source nature impacts Alibaba’s direct revenue from selling model APIs. There were also internal assessments of Qwen’s specific deliverables. Some senior executives were reportedly not fully satisfied with Qwen-3.5, which debuted on Chinese New Year’s Eve, describing it as a "half-finished product."
From Alibaba’s broader perspective, technological influence and open-source contributions are not end goals but means to achieve strategic and commercial objectives like AI cloud services and super AI apps. In the AI cloud segment, Alibaba Cloud faces aggressive competition from ByteDance’s Volcano Engine, which follows a closed-source model strategy. In the super app arena, the recent Spring Festival subsidy battle did not significantly narrow the gap between the Qianwen App and Doubao.
A deeper issue lies in the misalignment between commercial objectives and technical goals, and the tension between top-down strategic planning and the independent exploration of internal teams.
Lin Junyang, Yu Bowen, and Hui Binyuan—all of whom recently left—began their careers at Alibaba as fresh graduates nurtured by the company. Lin joined Alibaba DAMO Academy in 2019 after earning a master’s degree in Linguistics and Applied Linguistics from Peking University. Yu joined in 2022 after obtaining a Ph.D. from the Institute of Information Engineering, Chinese Academy of Sciences, and was recognized as an "Alibaba Star" that year. Hui, born in 1999, officially joined Alibaba DAMO Academy in 2022 after receiving a master’s degree from Tianjin University. All three participated in the early training of Qwen models.
Lin possesses a cross-disciplinary background in computational linguistics and AI. Not an overseas returnee, he is a locally trained AI technical leader with international influence. In 2025, at age 32, he became the youngest P10-level employee at Alibaba.
Zhou Hao, who recently joined Alibaba, earned his Ph.D. from the University of Wisconsin-Madison in 2019. According to his LinkedIn profile, he was a key contributor to projects like Gemini 3.0, Al Mode, DeepResearch, and Gemini 1.0, leading multi-step reinforcement learning efforts for Gemini 3.0.
Lin’s departure has sparked widespread discussion within the AI community, with many colleagues and industry professionals expressing regret and gratitude on social media. Those familiar with his management style describe him as supportive, focused on inspiring team autonomy and cohesion, and valuing logical reasoning. He believed that the most important task for any team lead was to hire people better than themselves, which required having a small ego and avoiding a sense of invincibility.
A Tongyi Lab insider previously noted that before the AI boom of 2023, the Qwen team developed largely unnoticed, allowing them to focus on model iteration with minimal disruption. However, as AI became an all-out battle that major tech companies cannot afford to lose, core model R&D teams began facing more organizational changes. While such changes often follow significant setbacks, Alibaba’s restructuring occurred at a time when external evaluations and internal morale were relatively high.
Lin’s resignation came as a surprise to Alibaba as well. For the company, individual needs must ultimately yield to organizational requirements.
Comments