(Yicai) March 6 -- The recent departure of Lin Junyang, the technical leader of Alibaba Group Holding’s Qwen team, is closely tied to the firm’s internal competition mechanism, which sets different teams to work on the same high-stakes project, a source close to the large language model developer, which is still facing challenges in the LLM arena, told Yicai.

Before leaving, Lin was engaged in an internal "competition" with Steven Hoi, a prominent artificial intelligence scholar who joined the Chinese e-commerce giant in February last year, the insider said. Zhou Hao, a former senior research scientist at Google’s DeepMind who was hired by Alibaba earlier this year, could be next in line as he is also entering the company’s internal competition system.

Hoi, formerly a tenured professor at Singapore Management University, joined Hangzhou-based Alibaba in February 2025 as chief scientist of Alibaba's Intelligent Information Business Group. His main responsibilities were multimodal foundational models and agent-related research for AI-to-Consumer applications. In September 2025, Hoi transitioned to the firm’s Tongyi Lab to focus on multimodal interaction models.

Both Hoi and Lin reported directly to Alibaba Cloud Chief Technology Officer Zhou Jingren, the insider said. However, Hoi's output in the internal race was reportedly not ideal, while Lin's Qwen team delivered the Qwen 3.5 series during the Chinese New Year break last month.

"Racehorse" is an internal competition mechanism adopted by several large internet companies. Simply put, when a company faces a new project with a high degree of uncertainty or of strategic importance, it allows different internal teams to start simultaneously and explore different approaches.

Zhou Hao is the next potential candidate for the internal race, the source said. Zhou Hao reports directly to Zhou Jingren, which intensifies internal competition by adding another peer-level leader and puts him in a similar situation to that of Hoi.

On the same day that Lin resigned, Yu Bowen, head of post-training for Qwen, also announced his departure. Additionally, Hui Binyuan, head of Qwen Code, had already joined Meta in January.

In an internal email dated March 5, Eddie Wu said that Alibaba would "continue to adhere to its open-source model strategy and increase efforts to recruit top talent."

Commercialization Challenges

Analysts suggested that Alibaba has not yet positioned itself securely in the LLM field, so strengthening talent density is essential. Although Qwen performs well in open-source ecosystem indicators, it faces challenges in terms of commercialization.

Since Qwen was first open sourced in April 2023, it has been downloaded more than one billion times globally, with over 200,000 derivative models, according to Hugging Face data. This makes it the world’s most downloaded open-source model and the model with the most derivatives, far ahead of DeepSeek's 50 million plus downloads and more than 2,000 derivatives.

However, when it comes to model usage, which is a more direct market indicator, Qwen is not in the lead. In the week Feb. 16 to Feb. 22, the top five models in OpenRouter’s weekly rankings were MiniMax M2.5, Moonshot Ai’s Kimi K2.5, Zhipu·AI’s GLM-5 and DeepSeek V3.2, while Qwen was absent. On the OpenClaw AI agent usage charts, K2.5 consistently holds the top position, and Qwen 3.5 has failed to break into the top ranks. Its massive open-source ecosystem has not yet translated into real-world usage advantages.

A source close to Alibaba Cloud told Yicai that Alibaba's ultimate view is that models alone are not enough to form a technological moat and they will ultimately become infrastructure. Therefore, models’ commercial value is more likely to come through cloud services, with the focus on selling computing power. Open-sourcing will help build global influence needed for this business model.

However, IDC data shows that in the first half of last year, China's AI cloud market was led by Volcano Engine with 49.2 percent market share, while Alibaba Cloud had only 27 percent. Despite Qwen’s vast open-source ecosystem, this has not translated into market leadership for Alibaba Cloud.

Editor: Kim Taylor