SYLIN

DFlash

by Z Lab · Model · Language & code models

DFlash 2 keeps the top 16 candidates per draft position, picks the best path with a lightweight selector, and adds a small convolution, gaining one more accepted token per pass for about 1% extra latency

AccessOpen codeNeeds~77 GBDownload1.6 GBLicenseMIT

Evidence

  1. Nov 14
    With its official 8-draft-token setting on SGLang, DFlash 2 used about 73 GB VRAM on an H100
  2. Sep 7
    On MI300X/MI355X in vLLM, DFlash reached up to 2.87x throughput on Gemma 4 26B-A4B versus 2.11–2.27x for EAGLE-3
  3. Sep 7
    DFlash gave 1.77–2.06x on Qwen3.6-35B-A3B and up to 2.68x on Kimi-K2.5 on AMD GPUs; native MTP beat it on Qwen3.5
  4. Aug 30
    DFlash2 support is awaited in Unsloth and LM Studio; reportedly could lift Mac throughput to around 30 tok/s
  5. Aug 19
    DFlash 2 speculative decoding layer for Qwen3.8-27B claimed to run nearly 3x faster with identical output
Show all 9
  1. Aug 19
    On an A100 with SGLang, Qwen3.8-27B ran at 59.1 tok/s with DFlash 2 versus 28.9 tok/s baseline, just over 2x, on five prompts
  2. Aug 11
    Meta ships a 1.6 GB DFlash drafter for Muse Glimmer that guesses 16 tokens ahead; the big model verifies, output identical to running alone
  3. Aug 11
    With the DFlash drafter, Muse Glimmer GGUF went from 31.5 to 53.17 tokens/s on an A6000, about 1.7x faster
  4. Jul 27
    Lucebox implements DFlash: draft model proposes up to 16 tokens, target verifies them in one forward pass

Cost to run

Alternatives

JetSpecShown finishing the same answer faster than DFlash and plain autoregressive decoding in the project's demo
DSpineOn one H100 with Qwen3.8-27B, DSpine hit ~150 tok/s vs ~129 for DFlash 2, about 16% faster
Made from