Qwen3.5 Core Devs Resign Suddenly

Post LinkedIn

🗾Read original on ITmedia AI+ (日本)

#resignation #development-team #strategyqwen3.5

💡Qwen3.5 devs quit post-launch—open-weight future at risk?

⚡ 30-Second TL;DR

What Changed

Core Qwen3.5 developers resigned right after small model series launch

Why It Matters

Resignation may slow Qwen iteration pace, unsettling users of this competitive open model. Alibaba's assurance could stabilize confidence, but practitioners should watch for talent shifts.

What To Do Next

Check Qwen GitHub repo for any immediate release delays or new contributors.

Who should care:Developers & AI Engineers

🧠 Deep Insight

Web-grounded analysis with 7 cited sources.

🔑 Enhanced Key Takeaways

•Lin Junyang, the technical lead of Qwen, announced his resignation on X on March 4, 2026, just one day after the Qwen3.5 small models launch, citing team restructuring that reduced his authority[1][2][3].
•Multiple other key members departed including Yu Bowen (post-training head), Hui Binyuan (Qwen Code lead, joined Meta in January 2026), and Lin Kaixin (contributor to Qwen3.5, VL, Coder)[1][2][4].
•Alibaba's Tongyi Lab is restructuring from a vertically integrated team to horizontally divided modules directly overseen by lab leadership, with Zhou Hao (ex-DeepMind, Gemini contributor) replacing Yu Bowen[1][2][6].
•Qwen models achieved over 700 million Hugging Face downloads and 180,000 community derivatives by January 2026, highlighting its open-source impact prior to the resignations[1].

🔮 Future ImplicationsAI analysis grounded in cited sources

Qwen's open-weight releases will continue uninterrupted

Alibaba publicly affirmed commitment to the open-weight strategy amid the resignations, as stated in responses to inquiries[original].

Post-training capabilities may strengthen under new leadership

Zhou Hao, a former DeepMind researcher and Gemini 3.0 contributor, was recruited to lead Qwen’s post-training RL research[1].

Community fine-tuning derivatives will exceed 200,000 by mid-2026

With 180,000 derivatives already created from nearly 400 open-sourced models by January 2026, ongoing releases sustain momentum despite team changes[1].