Model techniques map
Techniquesinference & servingdecoding strategy

specific method · filed under inference & serving

DFlash

A speculative-decoding drafter using a lightweight block-diffusion model to produce a draft block in one forward pass.

Also called Block-6 DFlash speculative decoding, block-6 DFlash for speculative decoding, DFlash speculative decoding, DFlash-style speculative decoding drafter, DFlash-style, 5-layer SWA MTP drafter (DFlash-style).

sources
6
models
4
labs adopt it
3
strongest
core

How sources treat it

One count per evidence span, weakest treatment to strongest.

optional 3default 1core 4

Documented in

Further reading

Picked by hand, not extracted: where to read more, not evidence for anything on this page.

Evidence

8 spans quoted from the sources, strongest treatment first.

5-layer SWA MTP drafter (DFlash-style). Predicts 7 subsequent tokens per forward pass for parallel verification.

coremodel architecturein MiMo-V2.6-Pro-RLXiaomi

Speculative decoding in MiMo-V2.6 uses a multi-token prediction (MTP) module following the block diffusion design of DFlash

coreinference servingin MiMo-V2.6Xiaomi

it predicts 7 subsequent tokens in a single forward pass for parallel verification by the backbone

coreinference servingin MiMo-V2.6Xiaomi

5-layer SWA MTP drafter (DFlash-style). Predicts 7 subsequent tokens per forward pass for parallel verification.

coreinference servingin MiMo-V2.6-Flash-RLXiaomi

RL rollouts use block-6 DFlash for speculative decoding by default, replacing the multi-token prediction (MTP-3) configuration inherited from SFT.

defaultinference servingin MiMo-V2.6Xiaomi

NVIDIA is also releasing a DFlash draft model, which can be measured against the others and may perform best for your workloads.

optionalinference servingin Nemotron 3.5 LightningNVIDIA

DFlash: A speculative-decoding drafter that uses a lightweight block-diffusion model to generate an entire draft block in one forward pass.

optionalinference servingin Nemotron 3.5 LightningNVIDIA

Speculative decoding: a trained DFlash draft model is available for lower-latency serving

optionalinference servingin Laguna S 2.1Poolside

Filed alongside

Other methods under inference & serving :: decoding strategy.