MiniMax-M2.7
Open-weight model for coding and agent tasks
MiniMax-M2.7 is a language model built by MiniMax that focuses on software engineering and autonomous agent work. Released with open weights, it uses self-evolution techniques to improve its own skills and supports multi-agent collaboration for complex productivity tasks.
- Achieved 66.6% medal rate on MLE Bench Lite
- Scored 56.22% on SWE-Pro benchmark
- Supports native Agent Teams for collaboration
- Weights available on Hugging Face and ModelScope
- Reduced incident recovery time to under three minutes
What it is
MiniMax-M2.7 is a language model developed by MiniMax. It is designed to handle complex productivity tasks and build agent harnesses. The model participates in its own evolution by updating memory and refining skills during development. It supports native agent teams for multi-agent collaboration. Users can access it via API or download the weights for local deployment.
How it came about
The model was developed using a cycle of self-evolution. An internal version optimized a programming scaffold over more than one hundred rounds. This process involved analyzing failures, modifying code, and running evaluations. The company states this method led to a thirty per cent performance improvement. The model was released with open weights on platforms like Hugging Face.
Inside the model
M2.7 builds dozens of complex skills for reinforcement learning experiments. It improves its learning process based on experiment results. The model supports dynamic tool search and complex skills. It is designed for system-level reasoning, such as correlating monitoring metrics and conducting trace analysis. Recommended inference frameworks include SGLang, vLLM, and Transformers.
What it can do
On MLE Bench Lite, the model achieved a sixty-six point six per cent medal rate. It scored fifty-six point two two per cent on SWE-Pro. For real-world engineering, it reached seventy-six point five on SWE Multilingual. The model handles Word, Excel, and PowerPoint editing. It achieved an ELO score of 1495 on GDPval-AA. On Terminal Bench 2, it scored fifty-seven point zero per cent.
How to use it
The model weights are available on Hugging Face and ModelScope. MiniMax provides an API and a token plan for cloud access. Local deployment guides are provided for SGLang, vLLM, and Transformers. NVIDIA NIM endpoints also host the model. The company recommends specific inference parameters for best performance.
Questions people ask about MiniMax-M2.7
What is MiniMax-M2.7 used for?
It is used for professional software engineering, complex productivity tasks, and building agent harnesses. It handles coding, bug troubleshooting, and multi-agent collaboration.
How can I run MiniMax-M2.7 locally?
You can download the weights from Hugging Face or ModelScope. The company recommends using SGLang, vLLM, or Transformers for local serving.
Did the model improve itself during training?
Yes. The company states the model initiated a cycle of self-evolution, optimizing a programming scaffold over 100 rounds to achieve a 30% performance gain.
Good reading and sources
There is no English Wikipedia article on MiniMax-M2.7 yet, so the facts on this page come from the sources listed above: huggingface.co/MiniMaxAI/MiniMax-M2.7. 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.