Qwen3.6 Adds Thinking Preservation Flag

This is less a flashy capability jump than a crucial systems fix: Qwen is making its reasoning mode behave like something agents can actually build on. For developers shipping multi-turn workflows, that matters more than another benchmark slide.

• –`preserve_thinking` keeps prior reasoning available instead of discarding it, which should reduce the cache invalidation pain people hit with the older Qwen template path.
• –The biggest win is in agent loops: planning, tool use, code edits, and verification steps can stay anchored to the model’s previous decisions instead of starting fresh every turn.
• –It should also cut redundant tokens in longer sessions, since the model no longer has to rebuild the same reasoning state repeatedly.
• –The catch is ecosystem support: model-side support only helps if your serving stack and client actually pass the new flag correctly.
• –For local runtimes and wrappers, this is now a compatibility checkpoint, not an optional tweak.