Modelpublic

mlx-community/Qwen3.8-27B-MTP-mxfp8

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

6likes1.2kdownloads

Model Card

Qwen3.8-27B-MTP-mxfp8

This repository contains the Multi-Token Prediction (MTP) drafter weights split from `Qwen/Qwen3.8-27B` and quantized with mlx_vlm.convert for mlx-vlm speculative decoding.

This is not a standalone model. Use it as the draft model alongside a compatible Qwen3.8 27B target checkpoint.

Use with mlx-vlm

bash

mlx_vlm generate \
  --model mlx-community/Qwen3.8-27B-mxfp8 \
  --draft-model mlx-community/Qwen3.8-27B-MTP-mxfp8 \
  --prompt "Write a quicksort in Python." \
  --max-tokens 256 \
  --enable-thinking

--draft-kind mtp is detected automatically from model_type.

Model details

—Model type: qwen3_5_mtp
—MTP block size: 3
—Target architecture: Qwen3.8 27B
—Precision: MLX MXFP8, group size 32
—Runtime: MLX / mlx-vlm
—Source revision: 1d4bf0f2ff6012fd82039f2fa52739d0dd7c60c0

The adapter holds only the MTP drafter weights; the target supplies the token embeddings and language-model head at runtime. Use an adapter and target derived from the same Qwen3.8 27B checkpoint. The upstream Apache 2.0 license and model limitations continue to apply.