README.md

---
license: other
library_name: transformers
pipeline_tag: text-generation
tags:
- dflash
- speculative-decoding
- block-diffusion
- draft-model
- efficiency
- minimax
- minimax_m2
- diffusion-language-model
---

# MiniMax-M2.7-DFlash

[**Paper**](https://arxiv.org/abs/2602.06036) | [**GitHub**](https://github.com/z-lab/dflash) | [**Blog**](https://z-lab.ai/projects/dflash/)

**DFlash** is a speculative decoding method that uses a lightweight **block diffusion** model to draft multiple tokens in parallel. This is the drafter model, which must be paired with [MiniMaxAI/MiniMax-M2.7](https://huggingface.co/MiniMaxAI/MiniMax-M2.7).

<div align="center">
  <img src="https://huggingface.co/z-lab/gemma-4-31B-it-DFlash/resolve/main/assets/dflash_system.png" alt="DFlash Architecture" width="85%">
</div>

## Quick Start

### Installation

vLLM:

Check out [vLLM issue #46105](https://github.com/vllm-project/vllm/issues/46105).

SGLang:

```bash
uv pip install "git+https://github.com/sgl-project/sglang.git#subdirectory=python"
```

### Launch Server

vLLM:

Check out [vLLM issue #46105](https://github.com/vllm-project/vllm/issues/46105).

SGLang:

```bash
python -m sglang.launch_server \
  --model-path MiniMaxAI/MiniMax-M2.7 \
  --tp-size 4 \
  --speculative-algorithm DFLASH \
  --speculative-draft-model-path z-lab/MiniMax-M2.7-DFlash \
  --attention-backend trtllm_mha \
  --speculative-draft-attention-backend fa4 \
  --mem-fraction-static 0.8 \
  --trust-remote-code \
  --host 0.0.0.0 \
  --port 30000
```

### Usage

For SGLang, use port `30000`.

```python
from openai import OpenAI

client = OpenAI(base_url="http://localhost:30000/v1", api_key="EMPTY")

response = client.chat.completions.create(
    model="MiniMaxAI/MiniMax-M2.7",
    messages=[{"role": "user", "content": "Write a quicksort in Python."}],
    max_tokens=4096,
    temperature=0.0,
    extra_body={"chat_template_kwargs": {"enable_thinking": True}},
)
print(response.choices[0].message.content)
```

## Benchmark Results

**Setup:** 4 NVIDIA B200 GPUs per server/run, SGLang, tensor parallel size 4, target attention backend `trtllm_mha`, draft attention backend `fa4`, thinking enabled, max output length 4096, greedy decoding. Concurrency 1 uses 128 prompts; concurrency 32 uses 1024 prompts.

### Throughput

_Generated tokens/sec_

**Block Size = 8**

| Task | Concurrency | **DFlash** |
|---|---:|---:|
| Math500 | 1 | **331.12** |
|  | 32 | **4422.52** |
| GSM8K | 1 | **304.07** |
|  | 32 | **4202.09** |
| HumanEval | 1 | **333.44** |
|  | 32 | **4394.23** |
| MT-Bench | 1 | **350.84** |
|  | 32 | **4549.75** |

### Acceptance Length

| Task | c1 | c32 |
|---|---:|---:|
| Math500 | 3.561 | 3.658 |
| GSM8K | 3.481 | 3.586 |
| HumanEval | 3.610 | 3.657 |
| MT-Bench | 3.550 | 3.624 |

## Acknowledgements

Special thanks to [David Wang](https://davidwa.ng/) for his outstanding engineering support on this project. We are also grateful to [Modal](https://modal.com/), [InnoMatrix](https://innomatrix.ai), and [Yotta Labs](https://www.yottalabs.ai/) for providing the compute resources used to train this draft model.

## Citation

If you find DFlash useful, please cite our work. To share feedback on DFlash or request new model support, please fill out this form: [DFlash Feedback](https://forms.gle/4YNwfqb4nJdqn6hq9).

```bibtex
@article{chen2026dflash,
  title   = {{DFlash: Block Diffusion for Flash Speculative Decoding}},
  author  = {Chen, Jian and Liang, Yesheng and Liu, Zhijian},
  journal = {arXiv preprint arXiv:2602.06036},
  year    = {2026}
}
```