Robots Atlas>ROBOTS ATLAS
DF
AI Models

DFlash2

2 (2B drafter)
A lossless block-diffusion drafter for speculative decoding (Inco AI). A 2B draft model predicts a whole token block at once, giving up to 3.43x speedup without changing the target model's output.
โœ“ Activeโœ“ Public accessโš– Open sourceSpecialized AI
Parameters
2B (drafter)
parameters
Release date
1 January 2026
Access:DownloadDeployment:๐Ÿ’ป Local

Overview

DFlash2 is an inference optimization technique โ€” a block-diffusion drafter for speculative decoding โ€” shipped as a lightweight 2B draft model (BF16). Instead of predicting tokens one at a time, DFlash2 predicts a whole block of tokens in a single pass and keeps the top candidates at every position, after which a lightweight selector traces one coherent path through them.

Its backbone uses two-tap dynamic convolutions and is designed to produce 7-8 draft tokens per verification step. The method is lossless: greedy output exactly matches the target model, and the target's distribution is preserved during sampling.

It was created by Inco AI (with mirror repositories such as z-lab). DFlash2 achieves a 3.43x speedup at concurrency 1 on GSM8K. The method is described in the DFlash paper (ICML 2026, Chen, Liang, Liu). It is released under the Apache 2.0 license and is applied to models such as Qwen3.8 and GLM-5.3-Flash.

Classification
Specialized AI
Access & deployment
Download
Local
Weights: Open source
Key parameters
๐Ÿงฉ Parameters: 2B (drafter)
๐Ÿ“ฅ Input: text

Technical specification

Parameters
2B (drafter)
parameters
License
Apache 2.0
Hardware requirements
Lightweight 2B draft model (BF16), paired with a target model in speculative decoding; 7-8 draft tokens per verification step.
Modalities
โฌ‡ Input
text
โฌ† Output
text

Capabilities and applications

Native model capabilities
Real-time inference
The model's ability to generate responses with very low latency (>1000 tokens/sec) on specialized inference hardware (e.g. Cerebras WSE), enabling interactive, turn-by-turn collaboration with a human.
Category: coding
Language modeling
Ability to predict subsequent tokens and generate coherent natural-language text based on the preceding context.
Category: language

Benchmark results

1 benchmark
GSM8K (przyspieszenie speculative decoding)
concurrency 1
3.43ร—
๐Ÿ“„ technical_report
Lossless speedup versus the target model's decoding.

Technical architecture

Core Architecture
SDSpeculative Decoding

Sources and related pages