Qwen3.6 Key Option

Qwen 3.6 has changed to keep the previous turn's reasoning intact via a new preserve_thinking flag.

• According to the model page guidance, you should use "preserve_thinking": True instead of chat_template_kwargs: {"preserve_thinking": False}.
• Regarding the KV cache invalidation issue that was a problem in previous versions, cache reuse is more stable because reasoning is appended rather than re-serialized.
• This is especially advantageous in agent / tool-calling scenarios. Since the model can reference its own prior thought process, decision consistency improves, and unnecessary re-reasoning is reduced, which also lowers token consumption.
• It's explained that this has the effect of optimizing KV cache utilization in both reasoning mode and non-reasoning mode.

A verification method is also presented.

1. Have the model think of two 20-digit random numbers, and tell it to output only one without using tools.
2. In the next turn, ask it to show the other number.
3. With preserve_thinking off, it loses the previous reasoning and cannot remember the second number, while with it on, it can reference the immediately prior thinking and state the second number.

The author summarizes that this behavior lets you confirm whether the setting is properly turned on.