News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
fnMiniMax-M1

MiniMax-M1

Open-weight reasoning model with hybrid attention and long context

MiniMax-M1 is a large-scale reasoning model released by MiniMax in June 2025. It uses a hybrid Mixture-of-Experts architecture with lightning attention to support one million token contexts. The model focuses on efficient test-time compute for complex software engineering and mathematical tasks.

Share on WhatsApp

  • Supports native context length of 1 million tokens
  • Uses hybrid MoE architecture with lightning attention
  • Trained on 512 H800 GPUs in three weeks
  • Features novel CISPO reinforcement learning algorithm
  • Outperforms DeepSeek-R1 in software engineering tasks

What it is

MiniMax-M1 is a large language model designed for complex reasoning tasks. Released by MiniMax in June 2025, it builds upon the earlier MiniMax-Text-01 architecture. The model features 456 billion total parameters, with 45.9 billion activated for each token processed. It supports an extensive context window of one million tokens, allowing it to handle very long inputs. The system is optimised for tasks requiring deep logical deduction, such as advanced mathematics and real-world software engineering.

How it came about

The model was developed by a large team at MiniMax, with authors listed alphabetically in the technical report. It evolved from the MiniMax-Text-01 base. The developers focused on improving the efficiency of reinforcement learning during training. They introduced two versions of the model: one with a 40,000 token thinking budget and another with 80,000. The 40K version represents an intermediate stage of the training process used to create the 80K model. The weights were made publicly available via Hugging Face in mid-2025.

Inside the model

MiniMax-M1 uses a hybrid Mixture-of-Experts (MoE) design combined with a lightning attention mechanism. This structure helps scale test-time compute efficiently, reducing the floating-point operations needed for long generations. Training relied on large-scale reinforcement learning across diverse problem sets, including sandboxed coding environments. The team created a new algorithm called CISPO, which clips importance sampling weights to improve learning stability. This approach allowed full training on 512 H800 graphics processing units to finish in just three weeks.

What it can do

The model performs strongly in software engineering, tool use, and long-context understanding. On the SWE-bench Verified task, the 80K version scored 56.0, surpassing several comparable open-weight models. It also showed high accuracy in mathematical reasoning, scoring 86.0 on AIME 2024. In long-context tests like OpenAI-MRCR at one million tokens, it achieved a score of 56.2. The system supports function calling, enabling it to interact with external tools in a structured format for agentic workflows.

How to use it

MiniMax has released the model weights openly under an open-weight licence. Users can download the 40K and 80K variants from Hugging Face. For deployment, the company recommends using vLLM version 0.9.2 or higher to ensure compatibility and performance. Alternatively, the model can be run using the Transformers library. A chatbot interface is also available online for general evaluation and testing purposes.

Questions people ask about MiniMax-M1

What is the context window size of MiniMax-M1?

The model natively supports a context length of one million tokens. This is eight times larger than the context size of DeepSeek R1, making it suitable for processing very long documents or codebases.

How many parameters does MiniMax-M1 have?

The model contains a total of 456 billion parameters. However, it uses a Mixture-of-Experts architecture, so only 45.9 billion parameters are activated for each token during inference.

What makes MiniMax-M1 efficient for reasoning?

It uses a lightning attention mechanism that reduces computational costs for long outputs. Additionally, the CISPO algorithm improves reinforcement learning efficiency, allowing faster training on complex tasks.

Can I run MiniMax-M1 locally?

Yes, the weights are open and available on Hugging Face. You can deploy it using vLLM or Transformers. The company suggests using vLLM version 0.9.2 or above for best performance.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on MiniMax-M1 yet, so the facts on this page come from the sources listed above: huggingface.co/MiniMaxAI/MiniMax-M1-80k, arxiv.org/abs/2506.13585. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.