Tongyi Qianwen and Huawei Pangu leaders leaving, a major personnel shake-up in the giant model industry

（Source: Li Binghao Guanlan）

△“The technical leader is caught between capital’s will and the squeeze of technical ideals.”

By Li Binghao

In China’s 2025 artificial intelligence arena, on one side is the riot of blooming flowers in the form of the “red packet battle” and the DAU (daily active user) myth; on the other is terrifying organizational turmoil. On June 23, 2025, a formal internal anti-corruption notice from ByteDance tore open the cracks behind the glossy image: Qiaomu, the core technology负责人 of the Doubao large model, was officially dismissed and had all year-end bonuses deducted due to an alleged conflict of interest with HRBP Cheng, and investigation fraud. Tens of millions of options were rendered worthless.

This is not an isolated case. Previously, the technology负责人 of Alibaba’s Tongyi Qianwen (Qwen), Lin Junyang, and the负责人 of Huawei’s PanGu large model, Wang Yunhe, also announced their resignations one after another. When top AI talents repeatedly leave the “AI empire” they once personally helped build, China’s large-model industry is undergoing a deep paradigm shift—from capital frenzy to organizational restructuring, from model showdowns to the on-the-ground deployment of agents (Agent).

Intense battles, the white-hot competition among giant companies’ large models

Personnel upheavals in large models, in essence, reflect the fierce battle among OpenAI, Google, and China’s “big-name firms.” In Silicon Valley, the contest between OpenAI’s o1 model and Google’s Gemini has entered a white-hot phase; domestically, ByteDance, Alibaba, and Tencent are pouring more than 8 billion yuan into red packet subsidies during the 2025 Spring Festival to capture the “AI entry point.”

In this environment of extreme competition, organizational pressure is pushed to the limit. The evaluation logic used by big firms for their large-model teams is rapidly shifting from “technological leadership” to “commercial retention.” Data shows that although Tongyi Qianwen’s daily downloads during the Spring Festival temporarily jumped by 125.66% and Doubao’s DAU surpassed 100 million, industry expert Guo Tao points out that users driven purely by incentives often see their 30-day retention rate fall below 5%.

When technology-intensive investment on the market side encounters the “receding tide of user freshness,” the technology负责人 finds themselves wedged between capital’s will and technical ideals. While the Qiaomu case at ByteDance involves private conduct and compliance issues, the high-pressure environment and conflicts of interest it reflects are precisely a symptom of organizational distortion under large-model commercialization anxiety. In high-intensity competition, talent is no longer an asset, but a “variable” that can flow at any time—due to strategic adjustments, KPI shortfalls, or compliance reviews.

Constraints of high pay, the paradox of high talent mobility

In the large-model domain, a salary in the tens of millions is already standard, but it has not brought the expected stability; instead, it intensifies the rhythm of “leave once the job is done.”

Technical people often have a strong “pioneer spirit.” Take the core member of DeepSeek, Luo Fuli, as an example: after leading the development of a landmark model, she chose to shift to Xiaomi. The same pattern has played out at Alibaba and Huawei as well. On March 4, Lin Junyang left a line on social media: “Goodbye, my beloved Qwen,” sparking sighs across the industry. As a key contributor who helped build the Qwen series from nothing to something, his departure is seen as a clash between technical idealism and commercialization metrics after Alibaba’s large-model division entered an “operational implementation phase.”

That same month, Huawei’s PanGu large-model负责人, Wang Yunhe, announced a farewell to the Noah’s Ark Laboratory, where he had spent nine years. The movement paths of these top talents are highly similar: they stay through the most difficult stage of model research and development (0 to 1), then choose to leave when the product enters large-scale commercialization and the messy details of operations (1 to N). High pay has become a “golden handcuff” for big firms to retain people, yet it can’t lock in talent’s hunger for the next technical revolution. For Lin Junyang and Wang Yunhe, perhaps the big firm’s hierarchical system and operations-driven orientation are no longer the best place to pursue AGI (general artificial intelligence).

Market iteration, from model worship to the agent era

The deeper driver behind personnel changes is the brutal iteration of market logic. In the long post that pointed out the truth after Lin Junyang’s resignation, it becomes clear that AI is shifting from “training models” to “training agents (Agent).”

In the past two years, the industry has fallen into “reasoning worship” and a “parameter competition.” However, the cold thinking behind the red packet battle tells us that what users need is not a “problem-solving machine” that can write poems, but an “action partner” that can help them book flights, plan trips, and handle complex workflows. Lin Junyang put it plainly: “A longer reasoning chain doesn’t mean a smarter model.”

At present, the AI market replicates extremely fast; the sudden, flash-in-the-pan explosion of popularity is often accompanied by standardized intent. Large models are becoming the foundation, while the real battlefield lies in “environmental interaction.” The arrival of the agent era means AI tools will no longer be a single chat box, but systems embedded into a variety of physical and digital scenarios. Huawei’s focus on the B-end sector for PanGu’s large models, and Tongyi Qianwen’s layout in the open-source ecosystem—their core logic all shifts toward “executable, deployable.”

This paradigm shift requires organizations to transform from a “technology-center model” to a “scenario-center model,” which will inevitably dismantle the existing leadership system that previously had pure technology R&D at its core.

Closing words

Qiaomu’s departure is a warning bell at the compliance bottom line; Lin Junyang and Wang Yunhe’s goodbyes are the backdrop of a hero’s turn. On the high-speed AI racecar in China’s large-model landscape, personnel fluctuations during the gear-change period are extremely rational market behavior.

Talent mobility does not represent failure; it is the reallocation of intellectual resources. When top talent overflows from big firms and flows into startups, vertical fields, or hardware vendors (like Xiaomi), only then do AI seeds truly begin to take root across all industries.

The exit of large-model负责人 marks that the AI industry has moved beyond the stage of a simple “tech show arena,” and started entering a more stringent, pragmatic, action-oriented agent era. AI that has begun to “come alive” in the real world is far more full of life than the parameters shown in PPT.

A wealth of information, precise analysis—find it all in the Sina Finance APP