Qwen3.8-27B-DFlash2-4bit

The DFlash2 block drafter for Qwen3.8-27B, incoai/Qwen3.8-27B-DFlash2 at revision 015e795645c74b1a0eeef3b570031fb62e769bc5, in MLX affine 4-bit, group size 64. It is not a standalone language model: it drafts for its target, whose embedding and lm_head it borrows.

How it was made

convert_drafter.py (in this repo) quantizes every projection of the BF16 checkpoint exactly as Splash's loader does when it loads that checkpoint (runtime/model/GgufPreparation.cpp quantizeGroup, MLX's affine rule as mlx.core.quantize rounds it), and keeps the norms, the convolutions' base kernels and the candidate selector's two codebooks in BF16, as Splash keeps them (runtime/model/DraftCheckpoint.cpp). So these are the weights Splash drafts with.

• Run on incoai/Qwen3.6-35B-A3B-DFlash2, the same script reproduces mlx-community/Qwen3.6-35B-A3B-DFlash2-4bit (converted from Splash's own draft images) byte for byte, all 208 tensors.
• Against seabit-ai/Qwen3.8-27B-DFlash2-4bit (mlx_vlm.convert), all 173 shared tensors are identical; that conversion also quantizes the two selector codebooks, which this one, like Splash, keeps in BF16.

config.json is the source's, plus the quantization entry and the drafter's sliding window in dflash_config (use_swa, swa_window_size, from the source's use_sliding_window and sliding_window).

• Source model.safetensors SHA-256: 67fc76d68dc5a9415511a4f394ef744d67510cd20e93b37cc2cc7d28e4bab65c
• This model.safetensors SHA-256: 35e046965b0fac6956dca948e98f5c762948ae970b772dd8b87341e32c74b2a1

Use

With scratchy, as the default drafter of Qwen3.8-27B (mlx-community/Qwen3.8-27B-4bit).

The drafter is incoai's work, under its Apache-2.0 license; this repository only converts it.