Can We Disable Thinking in Code Model After Plan Stage?
Separate the plan and code models and send only the plan output to the code model to avoid its internal thinking.
Configure your pipeline to route plan output directly to the code model without the thinking prompt, and test instruction adherence.
Summary
A user is exploring whether the internal “thinking” phase of the Qwen3.6‑35B A3B code model can be disabled during the initial hand‑off from the plan model (Qwen3.6‑27B) while preserving it for subsequent steps. The plan model generates a structured plan, and the code model is expected to follow that plan directly. The user suggests routing the plan output to the code model without the thinking prompt to improve instruction adherence. No concrete solution is provided, but the discussion hints at separating the models and adjusting the prompt to avoid the code model’s internal reasoning.
The article highlights that disabling thinking could reduce latency and improve determinism, but it may also affect the code model’s ability to handle new tools or information introduced after the plan stage. The conversation underscores the need for a clear pipeline design that balances plan fidelity with code model flexibility.
Developers can experiment by sending only the plan text to the code model and omitting the thinking instruction, then evaluating the code output quality and compliance with the plan.
Key changes
- Qwen3.6‑27B used for planning
- Qwen3.6‑35B A3B used for coding
- Goal to disable thinking in code model during hand‑off
- Potential approach: separate prompt or instruction
- No concrete solution provided
- Discussion on instruction following
- Impact on latency and determinism
- Need for clear pipeline design