MiniMax-M3.1-Flash-DFlash2
This repository contains the DFlash 2 draft model for
MiniMaxAI/MiniMax-M3.1-Flash.
It is not a standalone language model: it runs inside a speculative
decoding server and drafts tokens for the target model to verify.
DFlash 2 is a block-diffusion drafter for speculative decoding. It predicts a whole block of tokens in a single pass and keeps the top candidates at every position. A lightweight selector then traces one coherent path through them. Two-tap dynamic convolutions in the backbone keep the draft from decaying toward the end of the block. Decoding is lossless: greedy output matches the target model exactly, and sampling preserves its distribution.
Quick Start
Set up SGLang for the target model with the MiniMax-M3.1 cookbook, then launch it with these speculative decoding options:
--speculative-algorithm DFLASH \
--speculative-draft-model-path incoai/MiniMax-M3.1-Flash-DFlash2 \
--speculative-draft-model-quantization unquant \
--speculative-num-draft-tokens 8 \
--speculative-draft-attention-backend fa4
See the blog post for more details.
Evaluation
- Runtime: SGLang (MiniMax-M3.1 build) on two NVIDIA GB300 GPUs (TP2/EP2), with MXFP8 target weights and an NVFP4 sparse KV cache
- Speculation block size: 8 (7 draft tokens per verification step)
- Sampling: MiniMax's default parameters (temperature 1.0, top-p 0.95, no top-k), with adaptive thinking and the default reasoning effort
- Maximum new tokens: 4096
- Samples: 128 per task
- Prompts: benchmark formatting from
z-lab/dflash; MT-Bench uses the first turn only
We compare autoregressive decoding, MiniMax-M3.1-Flash's native DSpark drafter, and DFlash 2. Both speculative methods propose seven draft tokens per verification step.
Acceptance Length
Acceptance length is the per-request mean of completion tokens divided by verification steps at concurrency 1. Higher is better.
| Task | DSpark | DFlash 2 |
|---|---|---|
| GSM8K | 4.41 | 5.18 |
| MATH-500 | 3.85 | 4.77 |
| HumanEval | 3.00 | 3.92 |
| MBPP | 3.42 | 4.30 |
| MT-Bench | 3.08 | 3.67 |
Throughput
Throughput is total output tokens divided by end-to-end wall time.
Each cell shows output tok/s (speedup vs. autoregressive).
Concurrency 1
| Task | Autoregressive | DSpark | DFlash 2 |
|---|---|---|---|
| GSM8K | 123.6 | 341.2 (2.76×) | 435.9 (3.53×) |
| MATH-500 | 122.2 | 331.5 (2.71×) | 410.1 (3.36×) |
| HumanEval | 123.2 | 257.3 (2.09×) | 337.7 (2.74×) |
| MBPP | 124.5 | 295.7 (2.37×) | 382.4 (3.07×) |
| MT-Bench | 121.9 | 221.6 (1.82×) | 280.6 (2.30×) |
Concurrency 8
| Task | Autoregressive | DSpark | DFlash 2 |
|---|---|---|---|
| GSM8K | 712.2 | 1,463.1 (2.05×) | 1,524.8 (2.14×) |
| MATH-500 | 669.4 | 1,362.5 (2.04×) | 1,718.4 (2.57×) |
| HumanEval | 695.3 | 1,085.4 (1.56×) | 1,415.4 (2.04×) |
| MBPP | 646.4 | 1,199.8 (1.86×) | 1,478.2 (2.29×) |
| MT-Bench | 706.2 | 966.2 (1.37×) | 1,176.9 (1.67×) |
Concurrency 32
| Task | Autoregressive | DSpark | DFlash 2 |
|---|---|---|---|
| GSM8K | 1,109.5 | 2,248.1 (2.03×) | 2,687.2 (2.42×) |
| MATH-500 | 1,550.5 | 2,910.7 (1.88×) | 3,838.3 (2.48×) |
| HumanEval | 1,665.3 | 2,439.3 (1.46×) | 3,237.0 (1.94×) |
| MBPP | 1,406.1 | 2,436.7 (1.73×) | 3,176.1 (2.26×) |
| MT-Bench | 1,597.6 | 2,127.3 (1.33×) | 2,594.7 (1.62×) |
Citation
If you find DFlash 2 useful, please cite:
@misc{inco2026dflash2,
title = {{DFlash 2: Keep Drafting Parallel}},
author = {{Inco AI}},
year = {2026},
month = {August},
url = {https://inco.ai/blog/dflash2/}
}
Please also cite the original DFlash paper:
@inproceedings{chen2026dflash,
title = {{DFlash: Block Diffusion for Flash Speculative Decoding}},
author = {Chen, Jian and Liang, Yesheng and Liu, Zhijian},
booktitle = {International Conference on Machine Learning (ICML)},
year = {2026}
}