MiniMax-M1 - MiniMax's latest open-source inference model

What is MiniMax-M1?

MiniMax-M1 is the latest open-source inference model from the MiniMax team. Based on a hybrid expert architecture (MoE) combined with lightning attention, it boasts a total of 456 billion parameters, with each token activating 45.9 billion parameters. The model surpasses domestic closed-source models and approaches the most advanced overseas models, offering the industry's highest cost-performance ratio. MiniMax-M1 natively supports context lengths of up to 1 million tokens and provides two inference budget versions: 40 and 80K, suitable for handling long inputs and complex inference tasks. In benchmark tests, MiniMax-M1 outperforms open-source models like DeepSeek on multiple performance metrics, demonstrating excellent performance in tasks such as complex software engineering, long context understanding, and tool usage. The model's efficient computation and powerful inference capabilities make it a strong foundation for next-generation language model agents.

Main functions of MiniMax-M1

• Long context processingIt supports 1 million tokens as input and 80,000 tokens as output, making it suitable for handling long documents and complex reasoning tasks.
• Efficient ReasoningIt offers two inference budget versions: 40K and 80K, optimizing computing resources and reducing inference costs.
• Multi-domain task optimizationIt excels in tasks such as mathematical reasoning, software engineering, long context understanding, and tool use, and is adaptable to diverse application scenarios.
• Function callIt supports structured function calls, can identify and output external function call parameters, and facilitates interaction with external tools.

MiniMax-M1 Technical Principles

• Hybrid Expert Architecture (MoE)Based on a hybrid expert architecture, the model is divided into multiple expert modules, each responsible for handling a specific subtask or subset of data. Input data is dynamically allocated to different expert modules according to its characteristics, achieving efficient utilization of computing resources and parallel processing capabilities. Based on the MoE architecture, the model can maintain high computational performance under large-scale parameters while supporting more complex task processing.
• Lightning AttentionThe Lightning Attention mechanism optimizes the computation process, reduces redundant calculations, and significantly improves the efficiency of the attention module. Using a sparse attention pattern, it focuses only on the key parts of the input sequence, further reducing computational complexity. It supports efficient processing of long sequence data, allowing models to handle contexts of up to 1 million tokens.
• Large-scale reinforcement learning (RL) trainingThis paper proposes a novel RL algorithm, CISPO, which trains users to perform large-scale reinforcement learning and optimizes the model's output based on reward signals to improve performance on complex tasks. It improves training efficiency and model performance by sampling weights based on pruning importance instead of token updates. A hybrid attention design naturally enhances the efficiency of RL, addressing the unique challenges of scaling RL with hybrid architectures.

MiniMax-M1 performance

• Software engineering tasksOn the SWE-bench validation benchmark, MiniMax-M1-40k and MiniMax-M1-80k achieved scores of 55.6% and 56.0% respectively, slightly lower than DeepSeek-R1-0528's 57.6%, but significantly outperforming other open-source weight models.
• Long context understanding taskLeveraging a million-level context window, the M1 series excels in long context understanding tasks, outperforming all open-source weighted models, even surpassing OpenAI o3 and Claude 4 Opus, ranking second globally, only slightly behind Gemini 2.5 Pro.
• Tool usage scenariosIn the proxy tool use case (TAU-bench), MiniMax-M1-40k outperforms all open-source weight models, surpassing Gemini-2.5 Pro.

MiniMax-M1 project address

• GitHub repository:https://github.com/MiniMax-AI/MiniMax-M1
• HuggingFace model library:https://huggingface.co/collections/MiniMaxAI/minimax-m1
• Technical Papers:https://github.com/MiniMax-AI/MiniMax-M1/blob/main/MiniMax_M1_tech_report

MiniMax-M1 product pricing

• API call inference cost pricing:
  • 0-32k input length:
    • Input Cost0.8 yuan per million tokens.
    • Output cost8 yuan per million tokens.
  • 32k-128k input length:
    • Input Cost1.2 yuan per million tokens.
    • Output cost16 yuan per million tokens.
  • 128k-1M Input Length:
    • Input Cost2.4 yuan per million tokens.
    • Output cost24 yuan per million tokens.
• APP and WebUnlimited free use on the MiniMax app and web.

Application scenarios of MiniMax-M1

• Complex Software Engineering:It supports code generation, optimization, debugging, and documentation generation, helping developers quickly implement functional modules and improve development efficiency.
• Long text processingIt can generate long reports, academic papers, novels, etc., and supports long text analysis and multi-document summaries to meet diverse needs.
• Mathematical and Logical ReasoningIt can solve complex mathematical problems, such as competition math problems and mathematical modeling, handle logical reasoning tasks, and provide clear problem-solving approaches.
• Tool usage and interactionAs an intelligent assistant, it calls upon external tools to complete multi-step tasks, providing automated solutions and improving work efficiency.