Open Models Almanac
Methodsinference & servingdecoding strategy

general family · filed under inference & serving

Standalone parallel speculative drafters

Standalone parallel speculative drafters are separate draft models trained to generate speculative tokens for a main model during decoding.1 They complement the native MTP head and are released as separately trained checkpoints for parallel speculative decoding.2

How this definition was written

Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.

  1. we release LiLiCorr (Rusanovsky et al., 2026), a standalone drafter for Nemotron 3.5 Super that targets low-concurrency serving, where decoding is strongly memory-bandwidth bound — NVIDIA Nemotron 3.5 Super and Lightning Technical Report
  2. This release builds on our earlier parallel-drafter work for Nemotron 3.5 Lightning, for which we trained and released separate DFlash (Chen et al., 2026a) and DSpark (Cheng et al., 2026) checkpoints — NVIDIA Nemotron 3.5 Super and Lightning Technical Report

Also called Parallel speculative drafters, parallel drafters, standalone parallel drafters.

sources
2
models
2
lab adopt it
1
strongest
used

Documented in

Evidence

3 quotes from 2 sources, one card per model, each opening with its strongest.

How sources treat it
optional 1used 2

parallel drafters propose all draft tokens in a single forward pass whose latency is nearly independent of the draft length.

usedinference servingin Nemotron 3.5 SuperNVIDIA
2 more from this model

develop standalone parallel speculative drafters for low-concurrency, memory-bound inference

usedinference servingin Nemotron 3.5 Super and Nemotron 3.5 LightningNVIDIA

standalone parallel drafters target low-concurrency, memory-bound deployments

optionalinference servingin NVIDIA Nemotron 3.5 Super and NVIDIA Nemotron 3.5 LightningNVIDIA

parallel drafters propose all draft tokens in a single forward pass whose latency is nearly independent of the draft length.

usedinference servingin Nemotron 3.5 SuperNVIDIA
1 more from this model

develop standalone parallel speculative drafters for low-concurrency, memory-bound inference

usedinference servingin Nemotron 3.5 Super and Nemotron 3.5 LightningNVIDIA
↑