Open Models Almanac
Methodsinference & servingdecoding strategy

specific method · filed under inference & serving

DFlash

DFlash is a small drafter that fills a whole block of draft tokens in one forward pass, conditioned on the last accepted token and on hidden states taken from several layers of the target, which it injects as extra keys and values into every draft layer. Drafting cost no longer grows with the draft length, but each position is guessed without seeing the tokens drafted before it in the block, so acceptance falls along the block.

Written for the explainer Speculative decoding in production: drafters, acceptance and load.

Also called Block-6 DFlash speculative decoding, block-6 DFlash for speculative decoding, DFlash speculative decoding, DFlash speculative drafter, DFlash-style speculative decoding drafter, DFlash-style.

sources
6
models
5
labs adopt it
3
strongest
core

Documented in

Evidence

7 quotes from 6 sources, one card per model, each opening with its strongest.

How sources treat it
optional 4default 1core 2

MiMo-V2.6-Flash

Xiaomiadopted2 sources

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
2 more from this model

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

MiMo-V2.6-Pro

Xiaomiadopted1 source

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
1 more from this model

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

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
1 more from this model

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

Further reading

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

↑