---
title: "Z Lab's DFlash Pushes Qwen3.6-27B to 5x Faster With Diffusion Drafting"
subtitle: "Z Lab's DFlash uses a diffusion model as a speculative drafter for Qwen3-27B, hitting over 6x lossless speedup and outpacing EAGLE-3 by 2.5x"
url: "https://alphasignal.ai/news/z-lab-s-dflash-pushes-qwen3-6-27b-to-5x-faster-with-diffusion-drafting"
author: "AlphaSignal Newsroom"
author_url: "https://alphasignal.ai/editorial-team"
company: "AlphaSignal"
company_url: "https://alphasignal.ai/company/alphasignal"
published: "2026-04-23T10:15:06.000Z"
updated: "2026-06-29T07:53:09.622Z"
type: "Model"
topics:
  - "Infra"
  - "Llms"
subtopics:
  - "Inference Optimization"
  - "Model Serving"
upvotes: 274
source_url: "https://huggingface.co/z-lab/Qwen3.6-27B-DFlash"
comments: 0
truncated: true
---

# Z Lab's DFlash Pushes Qwen3.6-27B to 5x Faster With Diffusion Drafting

> Z Lab's DFlash uses a diffusion model as a speculative drafter for Qwen3-27B, hitting over 6x lossless speedup and outpacing EAGLE-3 by 2.5x

Speculative decoding has been one of the most effective tricks for making large language models faster without changing their outputs. The idea is simple: a small, cheap draft model guesses the next several tokens, and the big target model verifies all of them in one parallel pass. If the guesses are good, you get multiple tokens for the price of one verification step. The catch is that the draft model itself still generates tokens one at a time, which limits how fast you can push things. [DFlash](https://arxiv.org/abs/2602.06036), from Z Lab, breaks that ceiling by replacing the autoregressive drafter with a block diffusion model that predicts an entire chunk of tokens in a single forward pass.

The [Qwen3.6-27B-DFlash](https://huggingface.co/z-lab/Qwen3.6-27B-DFlash) model on Hugging Face is the drafter component built for Alibaba's Qwen3-27B. It pairs with the full `Qwen/Qwen3.6-27B` target model and is already compatible with both vLLM and SGLang. The paper was accepted at ICML 2026, and the technique has since been adopted in production: Xiaomi's MiMo v2.5-Pro-UltraSpeed uses DFlash to hit over 1,000 output tokens per second.

## Why autoregressive drafting hits a wall

Autoregressive LLMs require inherently sequential decoding, leading to high inference latency and poor GPU utilization. Speculative decoding mitigates this bottleneck by using a fast draft model whose outputs are verified in parallel by the target LLM, but existing methods still rely on autoregressive drafting, which remains sequential and limits practical speedups.

Many speculative decoding methods, like the EAGLE series and the native multi-token prediction (MTP) modules in recent models like Gemma 4 and DeepSeek-V4, still rely on sequential autoregression in the draft model. The draft model generates draft tokens one-by-one, a poor fit for modern hardware and a limit on achievable speedup. In practice, this caps real-world gains from methods like EAGLE-3 at around 2-3x over baseline.

Diffusion LLMs offer a promising alternative by enabling parallel generation, but current diffusion models typically underperform compared with autoregressive models. DFlash threads the needle: use diffusion only for drafting, where quality is guaranteed by the target model's verification step, and keep the autoregressive model in charge of final outputs.

## The two-part trick that makes it work

DFlash is a speculative decoding method that uses a small diffusion-LLM draft model to predict an entire block of tokens in a single forward pass, conditioned on hidden states from the target model. There are two distinct innovations here that each contribute independently to the speedup.

_Preview only. Full text requires an AlphaSignal Pro account._

## Sources

- [z-lab/Qwen3.6-27B-DFlash · Hugging Face](https://huggingface.co/z-lab/Qwen3.6-27B-DFlash)
- [DFlash: Block Diffusion for Flash Speculative Decoding (arXiv)](https://arxiv.org/html/2602.06036v1)
- [The next generation of speculative decoding: DFlash and Spec V2 – LMSYS](https://www.lmsys.org/blog/2026-06-15-next-generation-speculative-decoding-dflash-v2/)
- [DFlash Project Page – Z Lab](https://z-lab.ai/projects/dflash/)
- [DFlash – vLLM Speculators Docs](https://docs.vllm.ai/projects/speculators/en/latest/user_guide/algorithms/dflash/)