PSA: Qwen3.6 ships with preserve_thinking. Make sure you have it on.
Qwen3.6 adds a preserve_thinking flag to keep prior reasoning in context and address the KV cache invalidation issue seen with the Qwen3.5 template. The post cites the Qwen3.6-35B-A3B model page and gives a two-turn 20-digit-number test: with preserve_thinking on, the model can return the second number from its earlier reasoning. The practical point is cross-turn reasoning retention for agent and tool workflows; LM Studio does not support it yet, and an oMLX PR is open.
Why it matters: Qwen3.6 adds a preserve_thinking switch that keeps the previous turn's reasoning in context, fixing Qwen3.5's KV cache invalidation and lost multi-turn memory.