Qwen's Former Lead Critiques Hybrid Thinking, Backs AI Agents

Junyang Lin, the former technical lead for Alibaba's Qwen project, announced his departure on March 3, 2026, and is now pursuing independent research. Lin's recent talk, "Qwen: Towards a Generalist Model / Agent," and a subsequent detailed post outline his evolving perspective, particularly his critique of "hybrid thinking" in favor of AI agents. The talk provided an overview of the Qwen model family, including QwQ-32B, Qwen2.5-Max, Qwen3, Qwen2.5-VL, and Qwen2.5-Omni, presenting benchmark comparisons against models like DeepSeek-R1, Grok 3 Beta, Gemini 2.5 Pro, and OpenAI's o-series.

Lin elaborated on Qwen3's architecture, highlighting "hybrid thinking modes" which combine step-by-step reasoning with near-instantaneous responses, and introduced dynamic thinking budgets to control model reasoning. Qwen3 also significantly expanded multilingual support to 119 languages and dialects, up from 29. The Qwen3 family encompasses models ranging from 0.6 billion to 235 billion parameters, available in various quantized formats such as GGUF, GPTQ, AWQ, and MLX, all under the Apache 2.0 license. The presentation included demonstrations for web development and deep research tasks.

In the "Future work" section of his talk, Lin indicated a shift towards agents, listing further pretraining, reinforcement learning with environment feedback, longer context windows, and expanded modalities as key development areas. He concluded with the statement, "Training models -> training agents," a sentiment he later expanded upon in his independent research post. This transition signifies a move away from solely focusing on foundational model development towards building more autonomous and interactive AI systems.