Qwen3.5 no-think mode needs custom template

Runtime plumbing matters more here than model quality. Qwen's docs point to a custom llama.cpp template as the practical way to disable thinking, and the post's settings line up with that recipe. For local users, turning thinking off trims latency and avoids long reasoning traces when they just want direct answers, though the toggle still varies across frameworks.