Sampling parameters
Do you use standard sampling or what? because idk know why but qwen3.5 9b always has been fool for me even on a easy task, i tested with different sampling and no one of they worked properly
Do you use standard sampling or what? because idk know why but qwen3.5 9b always has been fool for me even on a easy task, i tested with different sampling and no one of they worked properly
Hi, I'm using the official recommended sampling parameters from Qwen's documentation (Thinking mode for general tasks):
temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=1.5
These work well for me. If Qwen3.5 9B feels underperforming, it could be a few things:
Make sure you're using the latest llama.cpp build — older versions may have compatibility issues with Qwen3.5 architecture
Check if your quantization level is too aggressive (Q2/Q3 can degrade reasoning noticeably on 9B)
Try switching between thinking mode and instruct mode（--reasoning off）— the recommended sampling differs between them. For non-thinking/instruct mode, try
temperature=0.7, top_p=0.8 instead
Hope this helps!