malekoo/Qwen3.8-27B-MLX-8bit

Qwen3.8-27B — MLX, 8-bit (group size 64, affine)

This model was converted to MLX format (8-bit (group size 64, affine) quantization) from Qwen/Qwen3.8-27B using mlx-lm. Refer to the original model card for more details on the model. Weights: ~29 GB (8.501 effective bits/weight) on disk (bf16 MLX artifact: ~54 GB; the original checkpoint is ~55 GB including the vision encoder and MTP drafter, which MLX conversions drop).

This conversion is text-only. mlx-lm's qwen3_5 implementation strips the base model's vision encoder and multi-token-prediction drafter; this artifact does not accept images or videos. For vision use, run the original checkpoint with Transformers, vLLM, or SGLang.

Quantization changes the numerics. The base card's benchmark numbers do not transfer to this artifact as-is. Measured on this build: wikitext-2 (test) perplexity 6.9447, vs 6.9352 for the bf16 conversion in the identical harness (+0.14% relative, effectively lossless; 4-bit: 7.0871) — disjoint 2048-token windows, no overlap, single token stream. Decode ~18 tok/s on an M5 Max MacBook Pro at 28.9 GB peak memory (4-bit: ~33 tok/s at 15.5 GB; bf16: ~10 tok/s at 54 GB).

Model description (inherited from the base card)

Qwen3.8-27B is the compact dense model of the Qwen3.8 generation, built on the Qwen3.5 architecture: 27B parameters, 64 layers in a 3:1 hybrid of Gated DeltaNet linear attention and gated full attention (24Q/4KV, head dim 256), vocab 248,320, context 262,144 native. Thinking mode on by default with reasoning_effort control (xhigh default / medium / low) and preserve_thinking. See the base model card for details and benchmarks.

Use with mlx-lm

pip install -U mlx-lm
mlx_lm.generate --model malekoo/Qwen3.8-27B-MLX-8bit --prompt "Explain KV caches briefly." --max-tokens 2048

OpenAI-compatible server: mlx_lm.server --model malekoo/Qwen3.8-27B-MLX-8bit --port 8080. Thinking control via chat_template_kwargs: enable_thinking, preserve_thinking, reasoning_effort (xhigh/medium/low).

Recommended sampling (from the base card)

• 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 (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

Conversion provenance

• Source: Qwen/Qwen3.8-27B (bf16 safetensors, released 2026-08-14)
• Quantized with mlx_lm.convert -q --q-bits 8 --q-group-size 64
• Toolchain: mlx 0.32.1.dev20260814+3d23f7d87, mlx-lm at ml-explore/mlx-lm main 254d153

License and attribution

Apache-2.0, inherited from the base model. Copyright the Qwen team; this repository is a quantized conversion and claims no additional rights.

Citation

@misc{qwen38,
    title = {{Qwen3.8-Max}: A New Bar for Coding and Cowork},
    url = {https://qwen.ai/blog?id=qwen3.8},
    author = {{Qwen Team}},
    month = {August},
    year = {2026}
}