Qwen3.5 4B overthinks simple hello

The weird part here isn’t the answer, it’s that the local stack is exposing the model’s scratchpad for a throwaway prompt. That’s normal for some Qwen/Ollama setups, but it’s still a UX trap if you expected a clean, chatty one-liner.

• –Ollama’s Qwen3.5 page explicitly tags the family with `thinking`, so verbose reasoning output is part of the intended behavior on thinking-capable variants.
• –For a 4B local model, decoding defaults matter a lot; if you leave the model too unconstrained, it can ramble even on trivial prompts.
• –If you want terse chat, use a non-thinking/instruct variant or tighten your generation settings instead of assuming the model is broken.
• –This is a good reminder that local LLMs often fail first at presentation, not capability: the model may be fine, but the wrapper is too permissive.