Lin Junyang, the former lead engineer of the original Alibaba Qwen large model technology, made his first public statement weeks after leaving the company on the evening of March 26, deeply analyzing the next evolution of large model technology.

Lin Junyang pointed out that the industry is undergoing a transition from "reasoning-based thinking" to "agent-based thinking (Agentic Thinking)." He believes that over the past year, the industry has been focused on how to make models "think longer," but the core of the future will be whether the model can think in order to "take action" and continuously refine its plans through interaction with the real world.

Reflecting on the Qwen Development Journey: The Pain of Forcing "Thinking" and "Instructions" to Merge

In his article, Lin Junyang openly shared the team's attempts and lessons learned in early 2025. At the time, the team was ambitious and tried to build a unified system that could adjust the level of reasoning based on the difficulty of the question.

However, practice proved that the significant differences in distribution between reasoning data and instruction data led to the model performing mediocrely in both areas after forced integration: it appeared redundant and lacked decisiveness when thinking, and was unreliable and costly when executing instructions. This insight explains why Qwen later shifted to independently releasing Instruct and Thinking versions, providing valuable engineering references for the industry.

A New Standard for "Good Thinking": Being Able to Support Effective Actions Is Key

According to Lin Junyang, the length of the reasoning chain does not directly equate to the intelligence of the model. Blindly pursuing long reasoning chains often wastes computing power. He predicts that the focus of future research and development will shift from simply training models to training the entire agent system of "model + environment."