Qwen-3.6-27B-Merlin-DFlash (fine-tuned DFlash drafter)

NOTICE: This is a derivative work of z-lab/Qwen3.6-27B-DFlash (MIT License, Copyright (c) 2026 Jian Chen, Yesheng Liang, Zhijian Liu / z-lab). The weights were modified by fine-tuning. See NOTICE and LICENSE. Not an official z-lab release.

A DFlash block-diffusion draft model for speculative decoding, fine-tuned to draft for the target model magicworld7/Qwen-3.6-27B-Merlin (revision 5c294550) on image-to-Three.js code generation.

Changes made to the original model

• Warm start: z-lab/Qwen3.6-27B-DFlash @ 0919688658996800f86b895034249700e9481106 (same architecture: 5 layers, hidden 5120, 32/8 heads, head_dim 128, target taps [1,16,31,46,61], block size 8; all draft layers sliding-window 2048).
• Fine-tuned 1 epoch (250 steps) with LR 6e-5 on ~2,000 prompts whose responses were generated by the Merlin target model.
• Weights only; no architecture change. Checkpoint epoch_1_step_250.

Result (held-out validation, same target model)

drafter	val accept length (block 8)
z-lab/Qwen3.6-27B-DFlash (original)	5.58
this model	6.16

Usage (SGLang)

--speculative-algorithm DFLASH --speculative-draft-model-path <this repo> --speculative-num-draft-tokens 8

Citation (original DFlash work)

@article{chen2026dflash,
  title   = {{DFlash: Block Diffusion for Flash Speculative Decoding}},
  author  = {Chen, Jian and Liang, Yesheng and Liu, Zhijian},
  journal = {arXiv preprint arXiv:2602.06036},
  year    = {2026}
}