Qwen3.6-27B MLX 3-bit Released

A model converted from Qwen/Qwen3.6-27B into MLX format has been released. mlx-lm 0.31.2 was used for the conversion.

The quantization scheme is as follows.

• Embedding layer: 5-bit, group_size=64
• Prediction layer: 5-bit, group_size=64
• All other layers: 3-bit, group_size=64

The author explained that this configuration aims to balance compression ratio and quality. The repository also includes usage instructions via pip install mlx-lm, along with example code using load() and generate().

Recommended generation parameters for LM Studio were also provided.

• Distinguishes between Thinking Preserve mode and Instruct Mode
• Sets different temperature, top_p, top_k, etc. for general tasks, coding, and reasoning tasks

Per the model card, it has 27B parameters, BF16/U32, Safetensors, and MLX format, targeting Mac users with limited memory.