Qwen3.5 27B overthinks simple greetings

This looks less like a broken model and more like a serving-stack problem: Qwen3.5 is built to think by default, so casual chat gets buried under internal deliberation if you do not explicitly turn that off.

• –Qwen's docs show a non-thinking mode, which means the fix is usually in the template or API settings, not the weights.
• –Ollama and LM Studio will differ mostly in how they expose those controls, so the local runtime choice matters.
• –For assistants meant to answer quick greetings and FAQs, defaulting to reasoning mode is a UX tax.
• –The upside is that the same behavior can be useful for harder prompts, so the trick is making reasoning opt-in instead of always-on.