Qwen3.8-27B-MLX-8bit

This model was converted to MLX format from Qwen/Qwen3.8-27B using mlx-vlm version 0.6.8, and quantized to 8 bits, so it runs on Apple silicon through mlx-vlm. Refer to the original model card for more details on the model itself.

Pair it with Qwen3.8-27B-MTP-MLX-8bit to get speculative decoding, because Qwen carries a Multi-Token Prediction head that the MLX conversion moves into a separate repository.

Use with mlx-vlm

pip install -U mlx-vlm

mlx_vlm generate \
  --model vvsotnikov/Qwen3.8-27B-MLX-8bit \
  --draft-model vvsotnikov/Qwen3.8-27B-MTP-MLX-8bit \
  --prompt "Write a quicksort in Python." \
  --max-tokens 256 --temperature 0.6 --enable-thinking

Drop --draft-model to decode without the speculator.

How this was produced

mlx_vlm convert --hf-path Qwen/Qwen3.8-27B \
  --mlx-path Qwen3.8-27B-MLX-8bit -q --q-bits 8 --q-group-size 64

The quantization is uniform affine at 8 bits with group size 64, and it covers 498 modules of the language model including embed_tokens, lm_head and the linear_attn projections. The vision tower stays dense in bfloat16, since the converter skips multimodal modules by default.

Verification

Every artifact in this release passed an automated check before it was published, and nothing became public until its remote checksums matched the local build.

Check	Result
Quantization	{bits: 8, group_size: 64, mode: affine}
Quantized modules	498 of 2180 tensors
Vision tower	dense, 333 tensors
MTP tensors	0, moved to the drafter repository
Acceptance with the 8-bit drafter	94.2% of drafted tokens accepted, 2.88 accepted tokens/round, over 69 rounds, 16.021 tok/s

License and attribution

The weights derive from Qwen/Qwen3.8-27B under Apache 2.0, so the original license and its terms carry over. Read the license itself before you use this model, and refer to the upstream model card for the model's capabilities and limits.