Alibaba’s Qwen team lost its technical lead and at least two other key members within 24 hours of releasing the Qwen3.5 small model series, a development that has rattled the global open source AI community and raised serious questions about where Alibaba’s AI strategy is headed.
Junyang “Justin” Lin, the technical architect who built Qwen from an internal research project into a platform with over 600 million downloads, announced his departure on X with a brief post: “me stepping down. bye my beloved qwen.” Staff research scientist Binyuan Hui and intern Kaixin Li also posted about their exits. None of the three disclosed whether the departures were voluntary or the circumstances behind them.
A Final Release Before the Exit
The timing is striking. The Qwen3.5 small model series, ranging from 0.8B to 9B parameters, had just shipped to public praise, including a post from Elon Musk citing its “impressive intelligence density.” The models use a Gated DeltaNet hybrid architecture, combining a 3:1 ratio of linear attention to full attention, which allows a 9B-parameter model to compete with significantly larger systems while maintaining a 262,000-token context window.
Practically, these models can run natively on laptops, smartphones, and even web browsers. Lin had long championed this “algorithm-hardware co-design” approach as a way to deliver high-capability AI without requiring massive compute infrastructure, a philosophy he outlined publicly at the January 2026 Tsinghua AI Summit.
For developers, Qwen3.5 was also framed as a step toward what the team called the “Agentic Inflection,” where models move beyond conversational tasks toward autonomous operation, navigating interfaces and executing complex code.
The Corporate Pivot
Alibaba recently consolidated its AI operations into a structure called the “Qwen C-end Business Group,” merging model research with consumer hardware teams focused on AI-integrated glasses and rings. The reorganization signals a shift from research output toward product revenue.
Hao Zhou, a veteran of Google DeepMind’s Gemini team, has reportedly been appointed to lead Qwen going forward. The transition from Lin to Zhou suggests a move from what sources describe as a “research-first” culture toward one driven by commercial metrics, specifically Alibaba Cloud daily active user growth.
That shift carries real consequences for the more than 90,000 enterprises currently deploying Qwen through DingTalk or Alibaba Cloud. Those organizations chose Qwen partly because it offered open weights under an Apache 2.0 license, giving them the performance of a frontier model with the transparency to inspect and modify it. Future flagship releases, including the rumored Qwen3.5-Max, may not carry the same terms.
A Pattern Playing Out Again
The situation echoes what unfolded at Meta after the release of Llama 4 last spring, which drew criticism and was followed by an internal reorganization. Meta subsequently brought in Scale AI co-founder Alexandr Wang while prominent researcher Yann LeCun departed. The pattern is becoming familiar: a research team produces open, widely adopted work, investor pressure mounts, leadership changes, and open access becomes a negotiating chip rather than a core commitment.
Industry analysts cited by InfoWorld warn that Alibaba’s trajectory mirrors that shift. The “open” in open-weight models tends to erode when quarterly revenue targets enter the picture.
- Qwen3.5 models span 0.8B to 9B parameters
- Context window: 262,000 tokens
- Total Qwen downloads to date: over 600 million
- Enterprises using Qwen via Alibaba Cloud or DingTalk: 90,000+
Whether Alibaba’s new leadership maintains the open source commitments that made Qwen a global reference point is now the central question for every developer and enterprise that built on top of it.
Photo by Google DeepMind on Pexels
This article is a curated summary based on third-party sources. Source: Read the original article
