Qwen3.8-27B just not work
Hi, I recently tested the Qwen3.8-27B model, but I've encountered significant issues that make it difficult to use in practice.
When I try to ask even simple, straightforward questions, the model often gets stuck (freezing/hanging) or produces repetitive responses. It feels like the generation quality and stability are severely degraded.
For comparison, I'm currently running the Qwen3.6-35B-A3B, and it works flawlessly. It handles simple prompts quickly and consistently. It's surprising to see such a performance regression with the newer 3.7-27B model.
Has anyone else experienced this? Any tips would be greatly appreciated.
Start with the recommended parameters from the Qwen team.
Thinking Mode: temperature=1.0, top_p=0.95, top_k=20, min_p=0.0, presence_penalty=0.0, repetition_penalty=1.0
Instruct (or non-thinking) mode: temperature=0.7, top_p=0.80, top_k=20, min_p=0.0, presence_penalty=1.5, repetition_penalty=1.0
works like a charm here
mlx_lm --version
0.31.3
