Alibaba's Former Qwen Lead Declares 'Failure' of Hybrid AI Thinking, Advocates for 'Agentic' Future
Junyang Lin, ex-technical lead of Alibaba's Qwen, asserts the 'failure' of hybrid thinking in large language models, signaling a pivotal shift towards 'agentic thinking' as AI's next core direction.
✨ This content was summarized and interpreted by AI; it may contain errors — please verify accuracy with the original sources. Learn more
Listen to this story
Junyang Lin, former technical lead of Alibaba's Qwen, has publicly declared the "failure" of hybrid thinking modes in large language models, marking a pivotal shift in the AI development paradigm. Lin, who departed Alibaba in March 2026, articulated this revised perspective in a recent talk and a 6,000-word essay titled "From 'Reasoning' Thinking to 'Agentic' Thinking."
While Qwen3 models prominently feature "hybrid thinking" with distinct "Thinking Mode" for complex tasks and "Non-Thinking Mode" for rapid responses, offering dynamic control, Lin now contends that the practical implementation of combining these modes has been largely unsuccessful. He identifies a fundamental conflict: instruction models are optimized for speed and concision, whereas reasoning models demand extensive token use for deep deliberation and accuracy. Attempting to merge these divergent training objectives without extreme precision can compromise both, potentially leading to incomplete mode separation where reasoning behaviors "leak" into non-thinking outputs.
This critique underpins his strong advocacy for "agentic thinking," which he posits as the next core direction for AI. Lin defines agentic thinking as models that "reason through action," continuously interacting with an environment and dynamically updating plans based on real-world feedback, moving beyond static internal reasoning. The core inquiry shifts from merely "how long a model can think" to "how effectively it can think to support action." This vision prioritizes robust environment design, reliable rollout infrastructure, and sophisticated multi-agent coordination. Indeed, Lin has since launched a new AI lab focused on world models and embodied intelligence, underscoring his commitment to this agent-centric future. His candid assessment from within a leading LLM project serves as a crucial industry warning: the path to advanced AI may lie not in hybridizing static reasoning, but in building truly adaptive and interactive agents.