Qwen3.5-9B hits thinking loops in Ollama

Qwen 3.5 9B’s hyper-efficient reasoning architecture is a double-edged sword: it punches way above its weight class but oscillates wildly without strict constraint parameters.

• –The issue is often tied to high temperatures (1.0+) in small quantized versions; lowering temperature to 0.7-0.8 typically stabilizes the internal monologue.
• –Ollama's native `--think=false` flag or the `/set nothink` command can force-disable the reasoning path to bypass the loop entirely.
• –System prompts that explicitly limit reasoning steps to a fixed number (e.g., "Analyze in max 3 steps") have proven effective at forcing termination.
• –With a 256K native context and GPQA scores topping GPT-4o, the model is clearly optimized for "thinking" which UI wrappers aren't yet perfectly tuned to handle.