Build1 publisher3 min readPublished
A diffusion drafter lost to Gemma's own Assistant model on a 12GB RTX 3060
DFlash uses a diffusion model to draft three tokens at once and claims to beat EAGLE-3. On one consumer GPU running llama.cpp on a JavaScript coding task, it trailed the drafter Google ships with Gemma-4-12B-it.
The Engineer · Build desk
What happened
- A dev.to author compared the DFlash diffusion drafter against Gemma-4-12B-it's own Assistant model on an RTX 3060 with 12GB of VRAM, running llama.cpp on the same JavaScript coding task as a previous test.
- The DFlash project claims support for a wide range of models and significantly better performance than EAGLE-3, the autoregressive draft model that preceded it.
- DFlash comes from Z-Lab at UC San Diego, led by assistant professor Zhijian Liu, who is also a research scientist at NVIDIA.
Compiled by The EngineerSomething wrong?How this is made
Why it matters
- constraint Every supported checkpoint needs its own trained drafter, 21 of them today, so a base-model upgrade stalls the speedup until a matching drafter is released.
- decision A Gemma-4 operator choosing between the shipped Assistant model and a third-party drafter now has one data point saying the first-party option is the one to beat.
- contradiction The vendor claim is measured against EAGLE-3 and this test was measured against Gemma's tuned drafter, so the negative result prices the claim for a Gemma user without contradicting the paper.
- cost On a 12GB card the drafter's weights come out of the same budget as the 12B main model, so the memory spent on speculation is memory not spent on context or quantization headroom.
The drafter does not process the prompt itself. The main model runs its forward pass, emits the predicted token at N+1, and the DFlash drafter takes the main model's KV data by injection into its own cache, runs a diffusion model and a token decoder, and produces N+2, N+3 and N+4 together [5]. The main model then checks the batch in one step, using causal attention to mask and computing the probabilities in parallel [6]. Three drafted tokens plus the one the main model produced anyway puts the ceiling at four tokens per verification step [2].
That ceiling only arrives when all three guesses survive verification. The speed comes from correct guesses [7], and a rejected token costs a drafter forward pass for nothing. So any published speedup transfers to your workload only if your prompt distribution yields the same acceptance rate as the one on the harness that produced the number. The dev.to run is a single JavaScript coding task on a single GPU under llama.cpp [1].
The two results were measured against different baselines. DFlash's published claim is broad model support and significantly outperforming EAGLE-3 [3]; the author of the dev.to piece measured against the Assistant model that ships for Gemma-4-12B-it, which the write-up calls a dedicated, well-optimized MTP model of its own [16]. Both things can hold at once: DFlash can beat EAGLE-3 in Z-Lab's harness and still trail the drafter Google tuned against its own weights. EAGLE-3, according to the same write-up, arrived in 2025 without escaping the autoregressive approach and without a dramatic speed improvement [14], which makes it a soft target compared with a first-party drafter.
Support comes one checkpoint at a time. Counting the published list gives 21 checkpoints across 10 model families, each needing its own dedicated DFlash model [1] [10]. DeepSeek's DSpark went the narrow way instead and supports only DeepSeek-V4 and DeepSeek-V4-Flash [12]. Each of the 21 has to be re-released when a base model moves. Qwen alone accounts for eleven of them [11].
The hardware sets the other constraint. A bolt-on drafter loads alongside the main model [10], so on a 3060 the Gemma-4-12B-it weights and the drafter weights share the same 12GB [3]. The author did not report tokens per second or quantization settings for either configuration.
DFlash comes out of Z-Lab at UC San Diego, a group led by Zhijian Liu, an assistant professor there who is also a research scientist at NVIDIA [8], and is described in the paper "DFlash: Block Diffusion for Flash Speculative Decoding" [9]. The diffusion-in-text lineage it draws on runs through Meta's LLaDa and KDDI ELYZA Lab's ELYZA-Diffusion-Instruct-1.0-Dream-7B [15], while speculative decoding itself goes back to a 2023 Google DeepMind paper [13]. For an operator running Gemma-4 on one consumer card, the shipped Assistant model is the incumbent, and on this task DFlash did not beat it [2].
What to watch
- Tokens-per-second figures and per-token acceptance rates from the same 3060 run, which would let the comparison be priced instead of ranked.
- Whether Z-Lab ships drafters for new Gemma-4 or Qwen checkpoints soon after the base weights land, or well after.
- Whether anyone reproduces the DFlash-versus-EAGLE-3 comparison outside Z-Lab's own harness.