This is a 4-bit affine MLX conversion of Qwen/Qwen3.8-27B for Apple silicon. The original tokenizer, chat template, and image/video processor files are included.

Details

Architecture: qwen3_5 / Qwen3_5ForConditionalGeneration
Quantization: 4-bit affine, group size 64
Native context length: 262,144 tokens
Weight size: approximately 15 GiB
Conversion code: erokhins/mlx-vlm, commit 77f16c25

Generate

bash

python -m mlx_vlm.generate \
  --model matvei-aleksandrovich/Qwen3.8-27B-MLX-4bit \
  --prompt "Explain linear attention in simple terms." \
  --max-tokens 256

For speculative decoding, use the matching Qwen3.8-27B-MTP-4bit drafter. A drafter from another Qwen version is not a supported substitute.

Conversion

bash

mlx_vlm.convert \
  --hf-path Qwen/Qwen3.8-27B \
  --mlx-path ./Qwen3.8-27B-MLX-4bit \
  --quantize --q-bits 4

This community conversion is not an official Qwen release. Validate quality, performance, and memory use on your own workload.