Qwen3.8-27B-DFlash2-MLX

The DFlash2 block drafter and the tokenizer for Qwen3.8-27B, and nothing else: no target weights. It is the sidecar chad attaches to any Qwen3.8-27B checkpoint it loads — in particular to Unsloth's GGUF files, which chad runs natively in its MLX engine and which download from Unsloth, not from here.

Use it

uvx chad-code --model ~/models/Qwen3.8-27B-UD-Q3_K_XL.gguf

Download the GGUF you want from unsloth/Qwen3.8-27B-GGUF. chad reads the file directly (no conversion), fetches this repository on first use (~1.2 GB), builds a small model directory around the file and decodes with the drafter attached. Nothing here needs to be downloaded by hand.

Rough guide for a 24 GB Mac, drafter attached (the governor sizes the window from the actual machine):

Unsloth file	weights	context
UD-Q2_K_XL	9.8 GB	~170k
UD-IQ3_XXS	10.9 GB	~138k
UD-IQ3_S	12.0 GB	~107k
UD-Q3_K_XL	13.2 GB	~74k

Contents

• dflash/ — the DFlash2 drafter: a 1.9B model that reads the target's residual stream at five tapped layers and proposes a whole block of tokens in one forward, quantized to 4-bit group-64 in MLX format. A drafter reads the residual stream, which the target's weight quantization only perturbs, so one drafter trained against the bf16 base serves every quant of this model. chad verifies each block in one batched target forward with exact rejection sampling: greedy output is token-identical to unspeculated decoding (to kernel rounding), sampled output keeps the model's distribution at any temperature. It is in a subdirectory on purpose: mlx-lm globs model*.safetensors at the repo root, so it is invisible to a base-model loader.
• tokenizer.json, tokenizer_config.json, chat_template.jinja, generation_config.json — Qwen3.8-27B's tokenizer and chat template. chad uses this template rather than the one embedded in a GGUF because this one renders reasoning_effort=medium unless told otherwise (the GGUF's defaults to xhigh), and a byte-stable prompt is what keeps chad's prefix cache warm.

Provenance and license

Apache-2.0. The drafter weights are converted from z-lab/Qwen3.8-27B-DFlash2; the method is Chen et al., DFlash: Block Diffusion for Flash Speculative Decoding (arXiv:2602.06036). Only the quantization and the MLX layout are ours; please honor the original release's license and citation terms. The tokenizer files are Qwen3.8-27B's (Alibaba Cloud, Apache-2.0). See NOTICE.txt.