Is the Qianwen core talent upheaval causing Alibaba's large model to falter?

The Lantern Festival just passed, and the Tongyi Qianwen team experienced a major shakeup among its core personnel: Technical Director Lin Junyang has resigned, along with three other key members: Qwen Post-Training Lead Yu Bowen, Qwen Code Lead Hui Binyuan, and Qwen 3.5 & VL & Coder core contributor Li Kaixin.

This is not an ordinary departure of a technical leader but a systemic conflict involving organizational structure, resource allocation, and open-source strategy. Biteye attempts to reconstruct the full picture of this personnel earthquake and raises a more fundamental question: in the AI era, how should major companies position their technological ideals?

1. Sudden Blood Loss: Collective Departure of Core Staff Less than 24 hours after the release of Qwen 3.5, a small model praised by Elon Musk for its “astonishing intelligence density,” Lin Junyang, the CTO of Tongyi Qianwen, posted a brief farewell on X early in the morning:

As of press time, this post has received over 11,000 likes, 4.5 million views, and the comment section is filled with heartbreak. Lin Junyang, the youngest P10 tech expert at Alibaba at age 32, left with a sense of regret. His background exemplifies a new generation of Chinese AI talent. Cross-disciplinary background: born in 1993, undergraduate in Computer Science at Peking University, but chose to pursue a master’s in Linguistics. Perhaps this unconventional experience for an AI elite gave him an intuitive edge in multimodal and semantic understanding. Alibaba leap: joined DAMO Academy in 2019, led the development of OFA and Chinese CLIP. Led Qwen: became head of Tongyi Qianwen in 2022, and at age 32, was promoted to Alibaba’s youngest P10 in 2025. He was not alone in leaving. Qwen Post-Training Lead Yu Bowen also resigned simultaneously. Hours later, Qwen Code Lead Hui Binyuan posted “me too” and changed his profile to “former Qwen.”

Data source: XHunt plugin A few hours later, core contributor Li Kaixin, involved in Qwen 3.5 & VL & Coder, also announced his departure and changed his Twitter profile to “Pre Qwen.”

Data source: XHunt plugin This star team, which created a global download count exceeding 1 billion, over 200,000 derivative models, and consistently ranked at the top of open-source large models, seems to be disintegrating at an observable speed.

2. Causes: The Battle Between Individuals and Major Companies in the AI Era A tweet from Qwen team member @cherry_cc12 reveals only a glimpse of this storm. As internal meeting information gradually leaks, we attempt to piece together the full picture of this collective exodus.

2.1 Organizational Dilemmas: From Special Forces to Assembly Line Speculation suggests that the original Qwen Lab was a sharp team of tech geeks, all versatile specialists. Lin Junyang was like a squad leader, leading the charge. However, rumors indicate that the Qwen team plans to split up, transforming from a vertically integrated system covering different training processes and modalities into separate horizontal teams focused on pre-training, post-training, text, and multimodal tasks. This is a typical approach of traditional internet companies. Alibaba might have thought: the early Qwen Lab was an internal incubator project. After a year, it’s time to scale up applications. How to improve efficiency? Break down each process into SOPs, optimize each step, and overall efficiency will improve. But this idea is outdated. Just look at OpenClaw, which single-handedly created a huge ecosystem—demonstrating that the game in the AI era has fundamentally changed.

2.2 Resource Dilemmas: Do They Exist or Not? On one hand, “Qwen is the company’s top priority,” on the other, Wu Ma says, “Resources are hard to satisfy everyone.” This contradiction resembles leadership making empty promises without real follow-through. What does it mean to say “Qwen is the top priority”? Or “We’ve done our best as Chinese CEOs”? Or that resource bottlenecks are just “communication flow issues”? Who are they fooling? There are only two possibilities: First: Top management doesn’t actually prioritize Qwen; investing in Qwen is just AI FOMO. Second: There are two factions within leadership—one values Qwen highly, the other doesn’t. The neglectful faction then starts to block resources. In short, some top executives only pay lip service. So, even the product line claimed to be of “highest priority” can’t secure basic resources.

2.3 The Battle Between Individuals and Platforms: Who Can Overrule the Organization? The most heartbreaking leaked information is HR’s statement: “Cannot elevate to divine status; the company cannot accept irrational demands and will not try to retain at all costs.” Is this right? The AI talent war has already gone into overdrive: in 2024, Qwen’s former technical lead Zhou Chang left to start his own venture, then quietly joined ByteDance’s Seed team, which offered a 4-2 level + eight-figure annual salary “sky-high offer.” In 2025, Meta offered a staggering $200 million compensation package to poach Pang Ruoming from Apple, including high-value stocks and milestone-linked incentives tied directly to technological breakthroughs. Does HR even do competitive intelligence? Or is this wrong? This statement seems to carry a hint of China’s ancient philosophy: individuals cannot override the organization.

2.4 Political Struggles: Who Are You to Whom? Internally, it’s said that “no political considerations were involved throughout,” but also that “it’s necessary to consider where to position Zhou Hao for maximum efficiency.” This is interesting—implying that Zhou Hao must be placed somewhere within the organization, just a matter of where. Anyone familiar with palace intrigues knows: who can get things done isn’t as important as who is obedient. A harsh workplace truth: for most managers, whether someone can solve real problems or threatens their position is equally important. In startups, you can aim high; in big companies, upper management’s sense of security may matter more than your ability. Think about it carefully.

2.5 Mismatch Between Open Source and Commercial Strategy Deeper tensions stem from the misalignment of open source and commercial paths. Qwen has established a huge reputation in the global open-source community—downloads, derivative models, international recognition are all high. But open source doesn’t necessarily translate into users and revenue. Now that Qwen is growing, the group naturally asks: I’ve invested so much, shouldn’t I get some return?

3. Reflection: The Dilemma of Big Tech in AI This incident at Alibaba is not surprising at all. Have you watched “The Annual Meeting Cannot Stop”? It’s based on Alibaba’s story. One classic line: “If you can’t solve the problem, just solve the person raising it.” Alibaba’s logic seems to be: no matter who leaves, Qwen will keep turning. That phrase—“What we’re doing is grand; over 100 people aren’t enough, we need expansion”—feels like Alibaba no longer truly understands AI, and AI no longer understands Alibaba. Even the Web3 space next door is amused. The internet era empowered individuals through platforms, emphasizing standardized, process-driven, replicable organizational structures. Individuals depend on platforms, which define the rules. The AI era is evolving into a landscape where super individuals hold super bargaining power, even reversing the platform’s role. When big companies try to manage AI creativity with internet-era organizational logic, conflict is almost inevitable. Behind these organizational struggles is a collective confusion about how to manage geniuses. When HR asks employees, “What do you think your value is?” those who can truly shape the future have already voted with their feet.