Qwen3.8-27B-DFlash2-MLXFast-Q4

Qwen3.8-27B DFlash2 — MLXFast affine-4 artifact

This repository contains a proposal-only, MLX affine-4 quantization of incoai/Qwen3.8-27B-DFlash2 for the Yukon MLXFast Qwen3.8 challenge. It is not a standalone language model. A Qwen3.8-27B target model must verify every proposed token.

The original DFlash2 checkpoint and its z-lab/Qwen3.8-27B-DFlash2 mirror contain the same model.safetensors blob. This derived artifact was made from that exact public blob at the following immutable revisions:

• incoai/Qwen3.8-27B-DFlash2@adde41d8fde3a75dc905a7df0bd5088d2a44b5a1
• z-lab/Qwen3.8-27B-DFlash2@ac04198556d7e8867853cbc356807b969f311b05

Quantization

• Format: MLX affine quantization
• Linear weights: 4 bits, group size 64
• Non-linear parameters and selector codebooks: BF16
• Tensor count: 175
• Quantized linear modules: 47
• model.safetensors bytes: 1,265,644,781
• model.safetensors SHA-256: 95cd528f14ced30c4cb4933c2ae233f822c20cc6c38f7502adccef359c0141ce

The converted tensor tree was strict-loaded into a faithful MLX DFlash2 implementation before publication. The Yukon submission independently pins this file by immutable Hub revision, exact byte count, and the challenge's documented single-file tree digest.

Architecture retained

The artifact preserves the released five-layer DFlash2 drafter, including:

• target features from layers 5, 19, 33, 47, and 61;
• parallel block drafting with the target embedding and output head;
• 2-tap dynamic grouped convolutions before and after attention and MLP;
• a top-16, rank-256 adjacent-candidate path selector;
• sliding-window grouped-query attention.

Every output token remains target-authoritative: the draft is only a proposal, and the challenge's trusted parent verifies the full target trajectory.

Upstream

• DFlash2 technical post
• DFlash implementation
• Original DFlash2 model card

This derived artifact is published under the upstream Apache-2.0 license.