MiniMax 发布 M3.1-Flash 多模态 MoE 模型，支持 1M token 上下文
MiniMaxAI/MiniMax-M3.1-Flash
MiniMax 发布原生多模态 MoE 模型 MiniMax-M3.1-Flash，支持文本、图像和视频输入，上下文窗口最高 1M tokens。该模型总参数约 428B、每 token 激活 23B，在 M3 基础上改用全模型稀疏注意力、量化感知训练和 DSpark 投机解码以提升推理效率。
原文给出了与 DeepSeek、Kimi、GLM、GPT-5.6 等模型在多项基准上的逐项对比，读者可据此判断其智能体与编码能力的位置。
MiniMax-M3.1-Flash is a native multimodal mixture-of-experts model for coding, agentic workflows, computer use, and cybersecurity. It supports text, image, and video inputs with a context window of up to 1M tokens.
Building on MiniMax-M3, M3.1-Flash improves task performance and inference efficiency with sparse attention throughout the model, quantization-aware training, and DSpark speculative decoding.
Highlights:
- Agentic performance: A generational leap in agentic tasks, coding, and cybersecurity — 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents' Last Exam, and 82.2 on CyberGym.
- Efficient by design: Sparse attention throughout, quantization-aware training, and DSpark speculative decoding improve inference efficiency.
Model at a Glance
- Architecture: Native multimodal mixture-of-experts (MoE)
- Parameters: Approximately 428B total; 23B activated per token
- Inputs / output: Text, image, and video inputs; text output
- Context window: Up to 1,000,000 tokens
- Attention: MiniMax Sparse Attention (MSA) throughout the model
- Speculative decoding: DSpark draft head included
Benchmarks
| Benchmark | MiniMax M3.1 Flash | DeepSeek V4.1 Flash-1017 | DeepSeek V4.1 Flash | Kimi K3 | GLM-5.3 | GPT-5.6 Sol | Opus 5 |
|---|---|---|---|---|---|---|---|
Automation-Bench v1.0.6 |
62.8 | 60.5 | 54.8 | 46.7 | 48.8 | 45.8 | 50.3 |
| BankerToolBench | 76.2 | — | 73.7 | 69.1 | 74.5 | 74.6 | 76.2 |
| GDP.pdf AA | 20.8 | — | 12.8 | 22.0 | 11.2 | 27.2 | 21.6 |
| Terminal Bench 4.0 | 38.1 | 42.3 | 31.2 | 12.6 | 37.9 | 39.9 | 51.8 |
| LOCA-Bench 256K | 60.3 | — | — | 49.7 | 50.3 | 60.0 | 53.7 |
Agents’ Last Exam |
29.6 | — | — | 28.3 | — | 30.6 | 32.2 |
| OSWorld 2.0 | 61.7 | — | 54.3 | 58.3 | — | 65.7 | 70.6 |
PostTrain Bench v1.1 |
39.3 | — | 30.5 | 36.6 | 37.5 | 36.2 | 40.4 |
| DeepSWE v1.1 | 73.7 | 77.0 | 74.2 | 67.5 | 66.9 | 73.0 | 74.0 |
| NL2Repo-Bench | 61.2 | 68.2 | 65.4 | 58.0 | 58.0 | 56.8 | 75.3 |
| SWE Refactor Bench | 23.0 | — | 11.8 | 19.5 | 10.4 | 28.5 | 31.0 |
CyberGym |
82.2 | — | — | 80.0 | 84.5 | 83.6 | — |
BabyVision (with tools) |
91.0 | 92.5 | 89.6 | 85.7 | — | 88.9 | 94.1 |
| HLE (without tools) | 43.4 | 42.6 | 39.2 | 43.5 | 42.3 | 49.5 | 55.6 |
| CharXiv RQ | 82.5 | — | 69.0 | 84.8 | — | 85.2 | 83.7 |
Evaluation Notes
- Automation-Bench v1.0.6: 600 publicly released tasks.
- BankerToolBench: OpenCode 1.17.18; MiniMax-M2.7 judge.
- GDP.pdf AA: Taken directly from Artificial Analysis.
- Terminal Bench 4.0: Codex 0.149.1.
- LOCA-Bench 256K: Official implementation.
- Agents’ Last Exam: Official Overall setting; full-set results from the official ALE leaderboard only. Blank = no official result (CLI-only excluded), not zero.
- OSWorld 2.0: 0808 Full release; highest reported per model at release, except DeepSeek V4.1 (in-house reproduction).
- PostTrain Bench v1.1: Claude Code 2.1.251, 512K context; otherwise official setup.
- DeepSWE v1.1: mini-SWE-agent 2.4.5, 512K context.
- NL2Repo-Bench: Claude Code 2.1.233, 512K context; reward hacking = 0; two inconsistent tasks corrected (scores use the corrected set).
- SWE Refactor Bench: Claude Code 2.1.220, 200K context; otherwise official setup.
- CyberGym: Claude Code 2.1.263, 1M context.
- BabyVision (with tools): Claude Code 2.1.285; min image side 1,500 px.
- HLE (without tools): Text-only subset (2,158 questions).
- CharXiv RQ: GPT-5.5 judge.
Results are for the official weights; third-party quantizations may differ.
What’s New in M3.1-Flash
M3.1-Flash retains M3’s core mixture-of-experts architecture, with three main changes:
- Sparse attention throughout: the first three full-attention layers are replaced with sparse attention.
- Quantization-aware training: the routed MoE experts and the attention paths are quantized during training, reducing memory footprint and improving serving efficiency.
- DSpark speculative decoding: DSpark replaces the earlier EAGLE-like MTP approach. The release includes the draft head for use with compatible serving implementations.
How to Use
Local Deployment
Download the model:
hf download MiniMaxAI/MiniMax-M3.1-Flash --local-dir MiniMax-M3.1-Flash
We recommend the following inference frameworks to serve the model:
- SGLang — see SGLang cookbook.
- vLLM — see vLLM recipe.
- TokenSpeed — see TokenSpeed recipe.
Additional integrations in progress: Transformers, KTransformers, unsloth, and ATOM.
Multi-Chip Deployment
FlagOS — see Multi-Chip Support and Downloads.
Inference Settings
Recommended sampling settings: temperature=1.0 and top_p=0.95.
For reasoning, set reasoning_effort in the chat template: low, medium, high, xhigh, or max.
The model supports a context window of up to 1M tokens. Usable context length and throughput depend on GPU memory and the serving configuration.
Contact Us
Contact us at [email protected].
来源：MiniMax 模型发布 · hf-mirror.com