Qwen3.6-35B-A3B loops, repeats in agents

The pattern looks less like a universal model failure and more like a bad fit between this MoE model, long-context agent orchestration, and the default decoding preset. Qwen’s own docs recommend higher-temperature thinking configs and note that presence_penalty tuning can reduce endless repetition, which matches the community reports here.

• –Qwen3.6-35B-A3B is a sparse MoE with 35B total and 3B active parameters, so it can be more sensitive to generation settings than the denser 27B variant.
• –Official guidance for thinking mode is much hotter than many local defaults: `temperature=1.0`, `top_p=0.95`, `presence_penalty=1.5`, `repetition_penalty=1.0`.
• –The thread’s replies suggest low-temperature presets can trigger loops, while moderate temperature increases help more than aggressive repetition penalties.
• –If this shows up mainly in Agent Zero, the wrapper likely needs model-specific presets rather than a one-size-fits-all local config.
• –For long-context agent runs, the practical fix is usually to adjust sampling first, then inspect the chat template, thinking-mode settings, and any speculative decoding or cache configuration.