Alibaba's large language model ambitions have been jolted by an unexpected leadership departure. In the early hours of March 4, the head of Alibaba Group's Qwen artificial intelligence model division, Lin Junyang, posted a brief message on X: "me stepping down. bye my beloved Qwen."
The announcement immediately reverberated across the open-source AI community.
The timing added to the shock. Just a day earlier, Lin had been celebrating the release of the Qwen 3.5 small-sized model with his team. His abrupt exit has therefore left the industry scrambling for answers — not only about his personal next move but also about the future direction of Alibaba's flagship large model initiative.
Credit: X
Qwen's chief architect
Lin has widely been regarded as the driving architect behind Qwen's development.
Born in 1993, he graduated from Peking University's computer science department and later obtained a master's degree in linguistics. He joined the DAMO Academy in 2019 and within six years rose to the elite P10 technical rank, becoming one of Alibaba's youngest top-tier engineers.
Beyond technical leadership, Lin also built Qwen's reputation in the global open-source ecosystem. He remained highly active in developer communities, frequently responding directly to technical questions and engaging with contributors.
His philosophy — "models are products" — emphasised that foundational models should be developed with a clear product mindset rather than treated solely as research artefacts. In China's rapidly evolving large-model landscape, Lin stood out as one of the few leaders able to bridge deep technical expertise with product and market awareness.
Restructuring inside Qwen
Industry sources suggest Lin's departure may be linked to internal restructuring within Alibaba Cloud.
The Qwen team operates under the Tongyi Lab overseen by Zhou Jingren, vice president of Alibaba Group and CTO of Alibaba Cloud. According to insiders, the organisation is shifting from a vertically integrated structure — covering pre-training, post-training, multimodal models and infrastructure — to a more horizontally divided structure consisting of multiple specialised sub-teams.
Such changes reportedly reduced Lin's managerial scope. The decentralised structure also contrasts with his earlier emphasis on tight coupling between pre-training and post-training processes.
As a result, observers believe internal differences may have emerged over the technical roadmap for Qwen.
According to industry insider @fxtrader on X, Alibaba Cloud has recently changed the evaluation metrics applied to the Qwen team, introducing consumer-style indicators such as daily active users (DAU) to assess the foundational model group.
Some insiders also claim that new leadership began managing parts of the team directly while bypassing Lin, effectively sidelining him. Several sources familiar with the matter have attributed the episode to internal politics.
Talent exits at Qwen
Lin's exit is not the first high-level departure within the Qwen ecosystem.
Over the past two years, several senior figures have left the project, including former technical leader Zhou Chang, speech team head Yan Zhijie, and multimodal leader Bo Liefeng. Some moved to ByteDance, while others joined Meta.
This time, Lin and post-training leader Yu Bowen both stepping down has renewed concerns about the team's stability. Although Alibaba has recruited Zhou Hao, who previously participated in Google's Gemini project, to assume certain responsibilities, the continued loss of core personnel remains closely watched by the market.
Alibaba's Qwen strategy
The personnel shake-up comes just as Alibaba attempts to consolidate its AI strategy.
Shortly before the leadership changes, the company announced it would unify its business-facing and consumer-facing large-model brands under the single Qwen name, aiming to concentrate resources and strengthen brand identity.
Earlier this year, during the Spring Festival subsidy campaign, the Qwen app's daily active users surged from 7.07 million to 73.52 million, sharply raising its market visibility and bringing it closer to ByteDance's rival chatbot Doubao.
Yet whether this "unified branding" reflects deeper integration beyond marketing remains uncertain.
From a broader strategic perspective, Alibaba faces mounting competition. ByteDance's cloud division Volcengine has been aggressively expanding in the AI cloud market, while in the consumer app arena Qwen still trails Doubao.
For Alibaba, whose core business remains cloud computing, the ultimate challenge is translating technological capabilities into commercial value through API-based monetisation.
Tensions inside Alibaba's AI organisation
The deeper tension may lie in how Qwen evolves within Alibaba's corporate structure.
In its early stages, the project developed in a relatively independent environment focused on rapid model iteration. Over time, however, alignment with the priorities of a large technology company inevitably brings resource reallocation and organisational constraints.
Such shifts may also weaken the integrated development path Lin had championed, potentially creating fractures between different functional teams.
Another factor attracting attention is Lin's background. Unlike many Chinese AI leaders who studied overseas, he represents a new generation of top domestic talent trained entirely within China.
With companies such as ByteDance, Tencent and Meta actively recruiting Chinese AI researchers, high-profile engineers now have increasingly broad career options. Lin's next destination is therefore being closely watched across the industry.
What lies ahead for Qwen
The key question now is whether Qwen will move toward deeper consolidation or drift toward fragmentation.
In the short term, the absence of unified technical leadership could weigh on team morale and innovation momentum. For Alibaba, however, the real challenge goes beyond the departure of a single P10 engineer.
It lies in striking a workable balance between technical direction, commercial monetisation and organisational governance.
With one of its central figures stepping aside, the course Qwen charts next may ultimately determine Alibaba's position in China's intensifying race to build competitive large language models.
Article edited by Joseph Chen
