Qwen’s Former Lead on What Hybrid Thinking Got Wrong — and Why He Now Backs Agents

Short summary

Qwen's former technical lead Junyang Lin discusses the limitations of hybrid thinking modes and dynamic thinking budgets, explaining why Qwen3 should pivot to agentic thinking. He covers why agentic RL infrastructure is harder to implement and vulnerable to reward hacking. This represents a strategic shift in model design philosophy.

• •Hybrid thinking modes with dynamic thinking budgets had fundamental limitations
• •Shift to agentic thinking offers a better path forward for model architecture
• •Agentic RL infrastructure presents implementation challenges and reward hacking risks

Generated with AI, which can make mistakes.