Qwen3.5-0.8B stumbles in long-think mode

This looks less like a “Qwen is broken” moment and more like a classic tiny-model + wrong inference settings failure mode.

• –The thread itself highlights missing generation context (tokens, sampling, template), which makes the result hard to interpret as a fair model test.
• –Qwen’s official Hugging Face docs warn Qwen3.5-0.8B can get stuck in thinking loops and may fail to terminate under some sampling setups.
• –Qwen3.5-0.8B is intended for lightweight prototyping, not robust long-chain reasoning under aggressive think settings.
• –For local runs, template correctness, thinking-mode controls, and stop/stream safeguards matter as much as raw model quality.