Open Models Almanac
Methodsinference & servingdecoding strategy

specific method · filed under inference & serving

DSpark

DSpark keeps a parallel block drafter and adds a Markov head, a low-rank bias on each position's logits from the token drafted just before it, plus a confidence head that predicts each position's chance of acceptance. A scheduler uses those estimates and a measured table of step times to choose how many positions to verify per sequence at the current load.

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

Also called DSpark speculative decoding, DSpark speculative-decoding module, DSpark speculative decoding architecture, DSpark speculative decoding drafter, DSpark drafter, DSpark speculative drafter.

sources
14
models
8
labs adopt it
4
strongest
core

Documented in

Evidence

14 quotes from 14 sources, one card per model, each opening with its strongest.

How sources treat it
optional 6used 5core 3

Mistral-Large-4

Mistral AIadopted4 sources

uses the model's built-in DSpark draft head for faster generation through speculative decoding.

coreinference servingin Mistral Large 4Mistral AI
3 more from this model

We also adapt the DSpark [81] architecture for speculative decoding

usedunclearin Mistral Large 4Mistral AI

The checkpoint ships with an embedded DSpark speculative decoding draft head

optionalinference servingin Mistral Large 4Mistral AI

Both checkpoints ship the DSpark drafter. The speculative depth is fixed at 3 by the checkpoint.

coreunclearin Mistral Large 4Mistral AI

DeepSeek-V4.1-Flash

DeepSeekadopted2 sources

We also introduce the DSpark ... speculative decoding architecture to improve decoding efficiency through semi-autoregressive draft generation and confidence-scheduled verification.

coreunclearin DeepSeek-V4.1-FlashDeepSeek
1 more from this model

DSpark speculative decoding (semi-autoregressive draft generation with confidence-scheduled verification)

usedinference servingin DeepSeek-V4.1-FlashDeepSeek

Beyond MTP, two draft models are provided with Nemotron 3.5 Lightning: DSpark, which is recommended for DGX Spark inference workloads and low concurrency data center workloads.

optionalinference servingin Nemotron 3.5 LightningNVIDIA
1 more from this model

DSpark: A semi-autoregressive speculative-decoding drafter that proposes a whole block of candidate tokens in a single forward pass from a parallel backbone.

optionalinference servingin Nemotron 3.5 LightningNVIDIA

Further reading

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

↑