153: Alibaba–Qwen Personnel Changes: Misreadings, Current Situation, Foreshadowing and What Comes Next | solo
Summary
- 林俊旸 initiated his departure; he was not told to leave, and Qwen’s split by modality and training stage was a major factor in his decision to leave. The change came just as Qwen 3.5 had finished training and the team was exhausted after sustained overtime. On the afternoon of March 3, he wrote in the DingTalk group that he “could not face everyone”; in the early hours of March 4, he posted, “Me setting down by my beloved queen.” Earlier, Qwen Code head 惠斌源 and post-training head 喻博文 had also left.
- The claim that Qwen App DAU was added to the model team’s KPIs has no factual or logical basis, and 周浩 was not parachuted in to replace 林俊旸. Qwen App sits within the Smart Information Business Group managed by 吴永明; 周浩 actually took over post-training from 喻博文 and reports directly to 周靖人.
- The real thread running through this crisis is infra: reinforcement learning turned an infrastructure shortfall into a question of organizational authority and accountability. Qwen’s RL post-training on infrastructure supported by Alibaba Cloud PAI was underwhelming; performance improved markedly after switching to ByteDance’s open-source veRL framework. When models reached the several-hundred-billion-parameter scale, the existing setup hit another bottleneck. 林俊旸 then pushed to build infra in-house and went over 周靖人’s head to 吴永明 for hiring support, laying the groundwork for the organizational changes that followed.
- Alibaba’s management framed the split as an expansion of investment in foundation models, while Qwen members experienced it as the dismantling of the core collaboration mechanism. Management said the form and pace of communication may not have been handled well enough and that “individuals cannot be put on a pedestal, nor retained at any cost”; after the March 4 all-hands, the direction of the reorganization had still not been formally decided, and the team was disappointed by senior management’s explanation.
- Talent poaching is now out in the open, but attrition remains a risk: recruiters, headhunters and investors at other companies are all approaching Qwen members. 林俊旸 also said, “The brothers in a bind should keep working as originally arranged—no problem,” but the team may be facing its biggest organizational crisis to date.
- Qwen’s open-source reputation is colliding head-on with Alibaba’s commercial objectives. Qwen gained influence through a range of model sizes, a developer ecosystem and multimodal foundation models, but Alibaba is also asking how efficiently open source converts into AI Cloud business, how it affects model API revenue and how complete Qwen 3.5 really is. As AI Cloud faces Volcano Engine’s aggressive catch-up and Qwen App chases Doubao, the small environment with “fewer interruptions and less tug-of-war” is becoming impossible to sustain.
Deep dive
1. Three popular explanations diverge from the known facts
曼祺 first clarified that 林俊旸 was not told to leave, but had “pursued the goal he wanted despite knowing it might be impossible, and pushed through to the final step,” ultimately suffering something like a crisis of faith; the decision was abrupt, leaving management little time to react. Earlier, Qwen Code head 惠斌源 and post-training head 喻博文 had also left.
On the afternoon of March 3, 林俊旸 announced his decision to leave in Qwen’s DingTalk group, saying he “could not face everyone”; in the early hours of March 4, he posted, “Me setting down by my beloved queen.” The entire sequence unfolded quickly.
There is currently no factual evidence that Qwen App DAU was added to the team’s performance assessment, and the claim does not fit the reporting structure: Qwen App belongs to the Smart Information Business Group managed by 吴永明, while the model team still has to deliver performance and efficiency and meet demand for open-source models across sizes and capabilities.
周浩, who joined in January 2026, was not 林俊旸’s replacement; he took over post-training from 喻博文 and reports directly to 周靖人. But most frontline members had not known he had joined or that he was taking over, so they nevertheless felt as if they had been “suddenly notified.”
2. The split struck at what made Qwen’s training efficient
Qwen previously housed pre-training, post-training, infra and different modalities in a single team, with core members such as 刘大一恒 and 喻博文 working together for years. The team believed that frequent communication, trust and a refusal to pass the buck when problems arose were important reasons for its high training efficiency and success rate.
In the proposed reorganization, speech may be folded into Bailing and text-to-image may be folded into Tongyi Wanxiang. After 喻博文’s departure, post-training is led by 周浩, who reports directly to 周靖人 and sits outside 林俊旸’s management. The future of the infra team Qwen began building in-house in mid-2025 remains unclear.
This would turn the organization from one where pre-training and post-training were tightly coupled and infra was vertically integrated with pre-training into parallel teams split by modality and training stage. The change came just as Qwen 3.5 finished training and the team, worn down by sustained overtime, was exhausted; it therefore became a major factor in 林俊旸’s departure.
At the March 4 all-hands, 吴永明, 周靖人 and 蒋芳 said the changes were not a retrenchment of Qwen but an expansion of investment in foundation models: “Making Qwen succeed is a group-wide matter”; bringing in talent inevitably entails changes to the team’s formation. Management said the form and pace of communication may not have been handled well enough, but did not spell out the specific organization plan going forward; procedurally, no formal decision had yet been made.
3. Reinforcement learning turned the infra shortfall into a conflict over authority and accountability
Alibaba had invested in large models as early as 2020, releasing M6 and PLUG in 2021. At the start of the 2023 boom, the M6 team and DAMO Academy’s NLP team competed head-to-head; the former delivered better results, after which resources were concentrated in the Intelligent Computing Lab led by 周靖人, gradually forming Tongyi Lab.
周畅 left in summer 2024, partly because his technical interests leaned toward multimodality and vision, and partly because he felt the team’s compensation did not match the market; the latter issue improved substantially after his departure.
After OpenAI released o1, Qwen stepped up RL post-training, but the initial results were poor. The team initially suspected the algorithm, then revised its diagnosis to infra: around the 2025 Spring Festival, performance improved markedly after switching to ByteDance’s open-source RL framework veRL, effectively locating the problem “by controlling variables.”
By mid-2025, as Qwen updated its next-generation core model and began training models at the several-hundred-billion-parameter scale, veRL plus localized PAI modifications still struggled to support larger-scale training. Qwen began briefing existing members and recruiting its own infra staff, but 周靖人 preferred to have PAI continue supporting the various Tongyi model teams and was not inclined toward this organizational change.
4. A skip-level push to build infra in-house foreshadowed the changes to come
林俊旸 believed the lack of independent infra would directly affect model-training performance, and ultimately went around 周靖人 to seek 吴永明’s support for hiring. According to 曼祺, this amounted to a skip-level escalation under company procedure and “laid the groundwork for some of the organizational adjustments and changes Qwen would later face.”
The Qwen team’s perception was that Alibaba Cloud’s training and inference infra worked better for external large-model vendors such as Kimi and Zhipu than its support for Qwen, the internal model team. At the all-hands, management explained the gap as “historical reasons.”
Asked whether 林俊旸 could stay, management replied that “we cannot put an individual on a pedestal, nor retain him at any cost.” After the meeting, the team remained disappointed by the explanations and statements, morale was relatively low, and 林俊旸’s staying or leaving became one of members’ biggest immediate concerns.
5. Open-source reputation must be reweighed against Alibaba’s commercial goals
The publicized changes quickly triggered a talent scramble: recruiters from other companies spent the night of March 3 contacting Qwen members; headhunters tried to reach 林俊旸, saying multiple companies had openings; investors looked for core people who might leave or start companies; and some technical leaders were already considering how to steer their reports toward new opportunities.
林俊旸 later posted on Moments: “The brothers in a bind should keep working as originally arranged—no problem.” 曼祺 stressed that this was not a fact but a subjective inference: he may not want his own departure to cause the team to fall apart, but much remains uncertain, and this could be Qwen’s biggest organizational and team crisis to date.
Qwen has a strong reputation in the global open-source community; in the US, the best-known Chinese models are DeepSeek and Qwen. Qwen’s models across sizes are popular with small and midsize startups; companies such as Cursor fine-tune on top of them, while many embodied-AI companies in China also choose its open-source multimodal models. Qwen Max, its largest flagship model today, is not open source; 曼祺 recalled that 林俊旸 once said he hoped to push for the release of a 1T-parameter model.
Inside Alibaba, the question is whether the open-source ecosystem converts directly enough into AI Cloud, and whether open source could affect revenue from direct model-API sales. Some technical executives were not fully satisfied with Qwen 3.5, unveiled on Lunar New Year’s Eve, calling it a “half-finished product.” Meanwhile, Alibaba Cloud is facing an aggressive catch-up push from Volcano Engine; in the recently concluded Lunar New Year subsidy battle, Qwen App also failed to materially narrow its gap with Doubao, even though Alibaba only began treating it again as a flagship AI application at the end of last year.
曼祺 ultimately traced the conflict to goal alignment across organizational layers: Qwen grew in “a corner where few people paid attention, with few interruptions and little tug-of-war,” but AI has become a full-scale war that major technology companies cannot afford to lose. How much individuality a large company can tolerate from a small team is a question of principle as well as choice; the outcome remains to be seen.