Qwen3.6 keeps CoT context intact

Small flag, big effect: this turns hidden reasoning from disposable scratchpad into stateful context for agentic work.

• –`--chat-template-kwargs '{"preserve_thinking": true}'` is the difference between “works in theory” and “actually behaves consistently” in local setups
• –Keeping thinking tokens around should reduce redundant re-reasoning across turns, which matters for long tool-using sessions
• –The tradeoff is obvious: more context retained means more tokens on the wire, but less thrash in repeated deliberation
• –This is a wrapper/inference-contract issue as much as a model issue; the model can support it, but your serving stack has to pass it through correctly
• –For agent builders, this is the kind of feature that quietly improves reliability more than benchmark numbers do