Qwen3.8-27B-oQ4e-mtp

Qwen3.8-27B-oQ4e-mtp

This is an MLX oQ4e quantization of Qwen/Qwen3.8-27B, produced with oMLX 0.5.7.

Quantization

• Quantizer: oMLX oQ4e
• Base precision target: 4-bit affine, group size 64
• Effective mixed-precision budget reported by oMLX: approximately 4.9 bpw
• Calibration: oQe imatrix-weighted affine quantization
• Non-quantized tensor dtype: bfloat16
• Native MTP weights and configuration retained
• Vision tower retained; this is not a text-only conversion
• Output: four safetensors shards, approximately 15.83 GB

oQ4e uses oMLX layer-sensitivity planning together with activation-importance calibration. Architecture protection rules leave the vision encoder unquantized and protect output-critical tensors.

Multimodal configuration note

The official base checkpoint includes both preprocessor_config.json and video_preprocessor_config.json. oMLX 0.5.7 copied the image preprocessor configuration during quantization but did not copy the separate video preprocessor sidecar. The original video_preprocessor_config.json from Qwen/Qwen3.8-27B was therefore added unchanged to this repository so that the official image/video processor metadata is preserved.

No vision weights were reconstructed or borrowed from another checkpoint.

Validation

The converted checkpoint was validated locally with oMLX 0.5.7:

• Loaded successfully as a VLM.
• Native Lightning MTP was selected and activated at runtime.
• Text generation completed successfully with MTP draft acceptance recorded by oMLX.
• An OpenAI-compatible multimodal API request correctly identified the main subject of a test image.
• The checkpoint index contains the retained language_model.mtp.* tensors and the full vision_tower.* tensor set.

Suggested oMLX settings

For thinking mode, the Qwen model card recommends:

• temperature=1.0
• top_p=0.95
• top_k=20
• min_p=0.0
• presence_penalty=0.0
• repetition_penalty=1.0

This conversion supports enable_thinking, preserve_thinking, and reasoning_effort (xhigh, medium, or low). Native Lightning MTP can be enabled in the oMLX model settings.

License and attribution

The base model is licensed under Apache 2.0. See the included LICENSE file and the official Qwen3.8-27B model card for details.