32‑year‑old star engineer leaves Alibaba (阿里): the clash between "top‑level design" and "low‑level code" in the AI era

Departure from Tongyi Lab

It has been reported that Lin Junyang (林俊旸), the 32‑year‑old technical lead of the Qwen team at Alibaba’s Tongyi Lab (通义实验室), posted "me stepping down. bye my beloved qwen" on social media in the early hours of March 4, 2026 and formally resigned. Reports say the immediate trigger was an organizational push to split the previously vertically integrated model group into horizontally divided units—pre‑training, fine‑tuning, text, and multimodal teams—which Lin opposed because he believes pre‑training, infrastructure and training must remain tightly coupled for technical quality.

Why this matters inside China’s AI labs

The dispute is more than a personnel story. Qwen reportedly had built its own infra team to break away from the generic, multi‑tenant cloud platform model and to pursue extreme efficiency and tight adaptation. When a team that treats a model as a single “work” is carved into assembly‑line roles, technical leaders can lose control over the model’s lifecycle and become cogs in a much larger machine. Similar exoduses have occurred before at Baidu (百度), which spawned Horizon (地平线) and Pony.ai (小马智行), and through departures from Tencent (腾讯) and ByteDance (字节跳动), reflecting a recurring tension between scientific autonomy and industrial management.

Commercial pressure and geopolitical context

Alibaba’s shift — reportedly evaluating Qwen more by AI‑cloud market share and super‑app user metrics than by pure technical impact — reflects wider commercial pressures. Chinese internet giants grew on refined division of labor, KPI‑driven execution and "race" mechanisms; now they must also convert AI research into cloud revenue and user time. Add geopolitics: US export controls on advanced chips and software have increased incentives for in‑house infra and localization, tightening the tradeoffs between chasing short‑term commercial returns and sustaining long‑horizon foundational research.

The bigger question

Can large Chinese tech firms balance industrialized, centralized R&D with the protected, exploratory spaces that produced early breakthroughs? Silicon Valley has wrestled with the same fault lines—OpenAI, DeepMind and Meta have all reorganized amid debates over commercialization and autonomy—so this is a global question with local consequences. For top‑tier engineers like Lin, the core calculus is whether they will have enough scope to realize technical visions, or whether “top‑level design” will be subordinated to the metrics of platform monetization.