Qwen3.5 Small Peers Hit Repetition Loops

Yes, this looks like a real small-model tax, not just a bad prompt. Low temperature can lock these models onto a high-probability phrase loop, and once they start echoing themselves, there often isn’t enough capacity left to recover cleanly.

• –Qwen’s official sampling guidance for text runs much hotter than `0.1-0.3` and leans on `presence_penalty`, which is a strong hint that ultra-low temperature is the wrong default for this family.
• –The Qwen3.5-0.8B model card explicitly notes it is more prone to thinking loops, so the behavior is not surprising even by the vendor’s own admission.
• –Meta positions Llama 3.2 1B as lightweight, multilingual, and on-device friendly, but that is a very different promise from “robust long-form generation under aggressive deterministic decoding.”
• –Gemma 3 1B is also framed as a portable model for constrained hardware, so the 1B tier is where you should expect brittleness first, especially in open-ended chat.
• –Practical fixes are usually boring but effective: raise temperature modestly, keep `top_p` reasonable, use `presence_penalty` or `frequency_penalty`, cap output length, add stop sequences, and stream so you can cut off runaways early.