MiniMaxAI/MiniMax-M1-40k-hf

This repository is primarily for the Transformers framework. If you're using other open-source frameworks, please use the alternative repository: MiniMax-M1-40k

MiniMax-M1

1. Model Overview

We introduce MiniMax-M1, the world's first open-weight, large-scale hybrid-attention reasoning model. MiniMax-M1 is powered by a hybrid Mixture-of-Experts (MoE) architecture combined with a lightning attention mechanism. The model is developed based on our previous MiniMax-Text-01 model, which contains a total of 456 billion parameters with 45.9 billion parameters activated per token. Consistent with MiniMax-Text-01, the M1 model natively supports a context length of 1 million tokens, 8x the context size of DeepSeek R1. Furthermore, the lightning attention mechanism in MiniMax-M1 enables efficient scaling of test-time compute – For example, compared to DeepSeek R1, M1 consumes 25% of the FLOPs at a generation length of 100K tokens. These properties make M1 particularly suitable for complex tasks that require processing long inputs and thinking extensively. MiniMax-M1 is trained using large-scale reinforcement learning (RL) on diverse problems ranging from traditional mathematical reasoning to sandbox-based, real-world software engineering environments. We develop an efficient RL scaling framework for M1 highlighting two perspectives: (1) We propose CISPO, a novel algorithm that clips importance sampling weights instead of token updates, which outperforms other competitive RL variants; (2) Our hybrid-attention design naturally enhances the efficiency of RL, where we address unique challenges when scaling RL with the hybrid architecture. We train two versions of MiniMax-M1 models with 40K and 80K thinking budgets respectively. Experiments on standard benchmarks show that our models outperform other strong open-weight models such as the original DeepSeek-R1 and Qwen3-235B, particularly on complex software engineering, tool using, and long context tasks. With efficient scaling of test-time compute, MiniMax-M1 serves as a strong foundation for next-generation language model agents to reason and tackle real-world challenges.

Benchmark performance comparison of leading commercial and open-weight models across competition-level mathematics, coding, software engineering, agentic tool use, and long-context understanding tasks. We use the MiniMax-M1-80k model here for MiniMax-M1.

2. Evaluation

Performance of MiniMax-M1 on core benchmarks.

| Category | Task | MiniMax-M1-80K | MiniMax-M1-40K | Qwen3-235B-A22B | DeepSeek-R1-0528 | DeepSeek-R1 | Seed-Thinking-v1.5 | Claude 4 Opus | Gemini 2.5 Pro (06-05) | OpenAI-o3 | |:---|:---|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:|:---:| | | *Extended Thinking* | *80K* | *40K* | *32k* | *64k* | *32k* | *32k* | *64k* | *64k* | *100k* | | *Mathematics* | AIME 2024 | 86.0 | 83.3 | 85.7 | 91.4 | 79.8 | 86.7 | 76.0 | 92.0 | 91.6 | | | AIME 2025 | 76.9 | 74.6 | 81.5 | 87.5 | 70.0 | 74.0 | 75.5 | 88.0 | 88.9 | | | MATH-500 | 96.8 | 96.0 | 96.2 | 98.0 | 97.3 | 96.7 | 98.2 | 98.8 | 98.1 | | *General Coding* | LiveCodeBench *(24/8~25/5)* | 65.0 | 62.3 | 65.9 | 73.1 | 55.9 | 67.5 | 56.6 | 77.1 | 75.8 | | | FullStackBench | 68.3 | 67.6 | 62.9 | 69.4 | 70.1 | 69.9 | 70.3 | -- | 69.3 | | *Reasoning & Knowledge*| GPQA Diamond | 70.0 | 69.2 | 71.1 | 81.0 | 71.5 | 77.3 | 79.6 | 86.4 | 83.3 | | | HLE *(no tools)* | 8.4\* | 7.2\* | 7.6\* | 17.7\* | 8.6\* | 8.2 | 10.7 | 21.6 | 20.3 | | | ZebraLogic | 86.8 | 80.1 | 80.3 | 95.1 | 78.7 | 84.4 | 95.1 | 91.6 | 95.8 | | | MMLU-Pro | 81.1 | 80.6 | 83.0 | 85.0 | 84.0 | 87.0 | 85.0 | 86.0 | 85.0 | | *Software Engineering*| SWE-bench Verified| 56.0 | 55.6 | 34.4 | 57.6 | 49.2 | 47.0 | 72.5 | 67.2 | 69.1 | | *Long Context* | OpenAI-MRCR *(128k)* | 73.4 | 76.1 | 27.7 | 51.5 | 35.8 | 54.3 | 48.9 | 76.8 | 56.5 | | | OpenAI-MRCR *(1M)* | 56.2 | 58.6 | -- | -- | -- | -- | -- | 58.8 | -- | | | LongBench-v2 | 61.5 | 61.0 | 50.1 | 52.1 | 58.3 | 52.5 | 55.6 | 65.0 | 58.8 | | *Agentic Tool Use*| TAU-bench *(airline)* | 62.0 | 60.0 | 34.7 | 53.5 | -- | 44.0 | 59.6 | 50.0 | 52.0 | | | TAU-bench *(retail)* | 63.5 | 67.8 | 58.6 | 63.9 | -- | 55.7 | 81.4 | 67.0 | 73.9 | | *Factuality* | SimpleQA | 18.5 | 17.9 | 11.0 | 27.8 | 30.1 | 12.9 | -- | 54.0 | 49.4 | | *General Assistant*| MultiChallenge | 44.7 | 44.7 | 40.0 | 45.0 | 40.7 | 43.0 | 45.8 | 51.8 | 56.5 |

\* conducted on the text-only HLE subset.

Our models are evaluated with temperature=1.0, top_p=0.95.

SWE-bench methodology

We report results derived from the Agentless scaffold. Departing from the original pipeline, our methodology employs a two-stage localization process (without any embedding-based retrieval mechanisms): initial coarse-grained file localization followed by fine-grained localization to specific files and code elements. The values for our models are calculated on the subset of n=486 verified tasks which work on our infrastructure. The excluded 14 test cases that were incompatible with our internal infrastructure are: "astropy__astropy-7606", "astropy__astropy-8707", "astropy__astropy-8872", "django__django-10097", "matplotlib__matplotlib-20488", "psf__requests-2317", "psf__requests-2931", "psf__requests-5414", "pylint-dev__pylint-6528", "pylint-dev__pylint-7277", "sphinx-doc__sphinx-10435", "sphinx-doc__sphinx-7985", "sphinx-doc__sphinx-8269", "sphinx-doc__sphinx-8475"

TAU-bench methodology

We evaluate TAU-Bench with GPT-4.1 as user model and without any custom tools. The maximum number of interaction steps is 40. Our general system prompt is:

- In each round, you need to carefully examine the tools provided to you to determine if any can be used.
- You must adhere to all of the policies. Pay attention to the details in the terms. Solutions for most situations can be found within the