Former Qwen Tech Staff: Large Models Evolving Into Action-Oriented Entities
Lin Junyang, former lead engineer of the original Alibaba Qwen large model technology, delivered his first public remarks weeks after departing the company on March 26, offering a deep analysis of the next evolution in large model technology.
Lin Junyang noted that the industry is shifting from "reasoning-based thinking" to "agent-based thinking (Agentic Thinking)." He argues that while the past year focused on extending how long models "think," the future core lies in whether models can think to "take action" and continuously refine plans through real-world interaction.
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 from early 2025. At that time, the team aimed to build a unified system capable of adjusting reasoning depth based on question difficulty.
However, practice showed that the significant distribution differences between reasoning and instruction data caused the model to perform mediocrly in both areas after forced integration: it appeared redundant and indecisive when thinking, and 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 model intelligence. Blindly pursuing long reasoning chains often wastes computing power. He predicts that future R&D will shift from training models alone to training the entire agent system of "model + environment."
Related article
ChatGPT Deep Integration Unlocks 300,000 Movies and TV Shows via Conversation
As streaming services saturate the market with endless content, helping users cut through the noise to find what they actually want has emerged as a critical industry hurdle.On April 9, major streaming providers officially announced the integration o
How to fix Core Web Vitals for better SEO ranking
Transform your video editing workflow with AI-powered transcriptsTable of Contents:IntroductionKey Advantages of Automated TranscriptionEnsuring Accuracy and Refining OutputLeveraging Search Capabilities in TranscriptsSeamless Import of Existing File
DingTalk Rebuilds Architecture With AliAI’s MuleRun, Launching Full-Agent Era
DingTalk’s new CEO and head of the Wukong Business Unit, Chen Yusen, revealed a major organizational restructuring in an internal memo issued this afternoon. By creating the Core Platform Business Department and fully merging the AI Agent team, DingT
Related Special Topic Recommendations
Comments (0)
0/500
Lin Junyang, former lead engineer of the original Alibaba Qwen large model technology, delivered his first public remarks weeks after departing the company on March 26, offering a deep analysis of the next evolution in large model technology.
Lin Junyang noted that the industry is shifting from "reasoning-based thinking" to "agent-based thinking (Agentic Thinking)." He argues that while the past year focused on extending how long models "think," the future core lies in whether models can think to "take action" and continuously refine plans through real-world interaction.
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 from early 2025. At that time, the team aimed to build a unified system capable of adjusting reasoning depth based on question difficulty.
However, practice showed that the significant distribution differences between reasoning and instruction data caused the model to perform mediocrly in both areas after forced integration: it appeared redundant and indecisive when thinking, and 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 model intelligence. Blindly pursuing long reasoning chains often wastes computing power. He predicts that future R&D will shift from training models alone to training the entire agent system of "model + environment."
ChatGPT Deep Integration Unlocks 300,000 Movies and TV Shows via Conversation
As streaming services saturate the market with endless content, helping users cut through the noise to find what they actually want has emerged as a critical industry hurdle.On April 9, major streaming providers officially announced the integration o
How to fix Core Web Vitals for better SEO ranking
Transform your video editing workflow with AI-powered transcriptsTable of Contents:IntroductionKey Advantages of Automated TranscriptionEnsuring Accuracy and Refining OutputLeveraging Search Capabilities in TranscriptsSeamless Import of Existing File
DingTalk Rebuilds Architecture With AliAI’s MuleRun, Launching Full-Agent Era
DingTalk’s new CEO and head of the Wukong Business Unit, Chen Yusen, revealed a major organizational restructuring in an internal memo issued this afternoon. By creating the Core Platform Business Department and fully merging the AI Agent team, DingT





Home






