adrianmurray/Qwen3.8-27B-MTP-MLX-6bit

Qwen3.8-27B native MTP MLX 6-bit

This artifact is the native multi-token prediction head extracted from the official Qwen/Qwen3.8-27B checkpoint and quantized to affine 6-bit MLX weights. It is bound to a matching target model and is used as a speculative draft by Qwen Prime Runtime and DFlash verification.

It is not a separately trained DFlash diffusion model. Speculative verification remains lossless with respect to the configured target decoding policy; speed depends on draft acceptance and workload.

Identity

• Source model: Qwen/Qwen3.8-27B
• Source revision: 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0
• Target model: Qwen/Qwen3.8-27B
• Quantization: affine 6-bit, group size 64
• MTP layers: one
• Weight SHA-256: 0af49c2c931f9f98c6beb2a85cae899a1c839647aae28b18e9187af53340fab2

qwen-prime-runtime doctor verifies the source/target binding and the complete weight hash before the artifact is used.

Reproduction

uv run python -m harness.trainer.export_qwen38_mtp \
  --source /path/to/Qwen3.8-27B \
  --output /path/to/Qwen3.8-27B-MTP-MLX-6bit \
  --source-revision 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0

License

The original model is Copyright 2026 Alibaba Cloud and is licensed under Apache License 2.0. This directory includes that license and a modification notice. This independent derivative is not endorsed by Alibaba Cloud or the Qwen team.