---
license: mit
language:
- en
- zh
library_name: transformers
pipeline_tag: text-generation
tags:
- text-generation
- agent
- long-context
- code
- mxfp4
- fp4-quantization
- dflash
- speculative-decoding
- block-diffusion
---
# MiMo-V2.5-Pro-FP4-DFlash
MiMo-V2.5-Pro-FP4-DFlash is the underlying model that powers MiMo-V2.5-Pro-UltraSpeed:
- An **FP4-quantized backbone** that applies MXFP4 quantization to the MoE experts while keeping the rest of the model at higher precision, shrinking model size and memory-bandwidth pressure with near-lossless quality.
- A **BF16 DFlash drafter** for block-diffusion speculative decoding, which proposes a whole block of tokens per forward pass and lets the backbone verify them in one step.
Together they cut both the per-parameter bit width and the number of backbone forward passes, the two dominant costs of trillion-parameter decoding.
## 1. Introduction
At the trillion-parameter (1T) scale, even 8-bit (FP8/INT8) inference carries severe memory-footprint and memory-bandwidth costs. Lowering the parameter bit width translates directly into faster decoding. We therefore adopt FP4 quantization and block-diffusion speculative decoding. Key features of this release:
- **Expert-Only FP4 Quantization**: A blanket FP4 cast over the whole model tends to degrade accuracy and generalization on complex reasoning and code. Given MiMo-V2.5-Pro's MoE architecture where experts hold the vast majority of parameters and tolerate quantization best, we quantize only the MoE experts to FP4 (MXFP4) and keep the other modules at their original precision. Through FP4 QAT, the model retains near-lossless capability while substantially reducing size and saturating hardware bandwidth.
- **DFlash Speculative Decoding**: A lightweight block-diffusion drafter fills an entire block of masked positions in a single forward pass, removing the serial draft autoregression bottleneck of conventional speculative decoding while the backbone's verification preserves output quality.
## 2. FP4 Quantization
We quantize only the MoE experts to MXFP4 (block size 32) and keep attention projections and other modules at higher precision (the attention `o_proj` of every layer is excluded from FP4). With FP4 QAT, quality stays close to the FP8 baseline:
| Benchmark | MiMo-V2.5-Pro-FP8 | MiMo-V2.5-Pro-MXFP4 | Δ |
| :--- | :---: | :---: | :---: |
| **General Agent** | | | |
| Claw-Eval (pass^3) | 63.8 | 67.8 | +6.27% |
| Humanity's Last Exam | 48.0 | 47.0 | -2.08% |
| Humanity's Last Exam (without tool) | 34.0 | 33.0 | -2.94% |
| **Code Agent** | | | |
| SWE-Bench Pro | 57.2 | 58.8 | +2.80% |
| SWE-bench Verified | 78.9 | 77.4 | -1.90% |
## 3. Block-Diffusion Speculative Decoding (DFlash)
Conventional speculative decoding relies on a small draft model to guess the next tokens, which the large model then verifies; the rejection-sampling verification keeps the output lossless. Its bottleneck is that draft quality bounds the acceptance rate, while a stronger draft costs more compute.
To break this trade-off we adopt the block-level masked parallel-prediction approach DFlash: the draft fills an entire block of masked positions in one forward pass. We landed this on MiMo-V2.5-Pro with custom optimizations for trillion-scale MoE and long-context serving, using the Muon second-order optimizer and model self-distillation so that even a small mask block keeps a strong acceptance rate while pushing the draft-stage cost close to its limit:
- The drafter uses Sliding Window Attention (SWA) throughout, naturally aligned with the SWA design of the MiMo-V2 series. The draft no longer depends on the full prefix, so the per-prediction compute moves from linear-in-context-length to constant.
- During training the mask signal is sampled on the local GPU shard, so a single sequence yields tens of thousands of independent training signals covering positions at different context lengths in one step, aligning with the MiMo-V2 series' long-context capability while avoiding cross-device communication overhead.
In practice, we further cap the mask block size at 8 to lower verification overhead and raise concurrency.
| Scenario | Acceptance Length |
| :--- | :---: |
| WebDev | 6.30 |
| Math500 | 5.56 |
| HumanEval | 4.54 |
| MT-Bench | 3.18 |
| SWE-Bench | 4.29 |
## 4. Model Summary
| Component | Backbone | DFlash Drafter |
| :--- | :---: | :---: |
| **Architecture** | MiMoV2ForCausalLM | DFlashDraftModel |
| **Total / Active Params** | 1.02T / 42B | 5-layer draft |
| **Hidden Size** | 6144 | 6144 |
| **Num Layers** | 70 | 5 |
| **Num Attention Heads** | 128 | 128 |
| **Num KV Heads** | 8 (GQA) | 8 (GQA) |
| **Head Dim (QK / V)** | 192 / 128 | 128 / 128 |
| **SWA Window Size** | 128 | 1024 |
| **Block Size** | — | 8 |
| **Captured Backbone Layers** | — | [0, 15, 31, 47, 69] |
| **Backbone RoPE Base** | 5,000,000 | 5,000,000 |
| **Precision** | MXFP4 (experts) Mixed | BF16 |
| **Max Context Length** | 1M | — |
## 5. Deployment
DFlash inference with the FP4 backbone is supported in SGLang. The drafter is launched alongside the backbone via the speculative-decoding flags and inherits the backbone's tensor/expert-parallel topology.
### SGLang Deployment
The following is an example of running the model with SGLang. Point `--model` at this repository and `--speculative-draft-model-path` at its `dflash/` subdirectory.
```bash
python3 -m sglang.launch_server \
--model MiMo-V2.5-Pro-FP4-DFlash \
--speculative-algorithm DFLASH \
--speculative-draft-model-path MiMo-V2.5-Pro-FP4-DFlash/dflash \
--speculative-num-draft-tokens 8 \
--ep-size 16 \
--tensor-parallel-size 16 \
--data-parallel-size 2 \
--enable-dp-attention \
--enable-dp-lm-head \
--quantization fp8 \
--attention-backend fa3 \
--moe-dense-tp-size 1 \
--dtype bfloat16 \
--mem-fraction-static 0.65 \
--context-length 65536 \
--page-size 1 \
--trust-remote-code \
--disable-overlap-schedule \
--skip-server-warmup \
--dist-init-addr ${MASTER_ADDR}:20000 \
--nnodes ${WORLD_SIZE} \
--node-rank ${RANK} \
--host 0.0.0.0 \
--port 29999
```
## Citation
```bibtex
@misc{mimo2026v25pro_fp4dflash,
title={MiMo-V2.5-Pro-FP4-DFlash},
author={{Xiaomi MiMo Team}},
year={2026},
howpublished={\url{https://huggingface.co/collections/XiaomiMiMo/mimo-v25}},
}
```
## Contact
For questions or feedback, reach us at [mimo@xiaomi.com](mailto:mimo@xiaomi.com) or join our community:
- [WeChat Group](https://work.weixin.qq.com/apph5/external_room/join/group_mng?plg_id=c417f99bd9014b5dd894daa8bfe19790&)
- [Discord](https://discord.gg/WX2R2uNp)
- [Telegram](https://t.me/+3T-I0pekOVIyNDBl)
- [Reddit](https://www.reddit.com/r/XiaomiMiMo_Official/)
