MiniMax-M1-80k

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.

2. Evaluation

Performance of MiniMax-M1 on core benchmarks.

* 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 these policies.

3. Recommendations for Minimax-M1 Model Usage

To achieve the best results with the Minimax-M1 model, we suggest focusing on two key points: Inference Parameters and the System Prompt.

3.1. Inference Parameters

• Temperature: 1.0
• Top_p: 0.95

This setting is optimal for encouraging creativity and diversity in