MiniMax M2.7: The 230-Billion-Parameter Open-Source AI Model with Self-Evolution

MiniMax has released M2.7, a 230-billion-parameter AI model with self-evolution capabilities, available on Hugging Face.

MiniMax M2.7: The 230-Billion-Parameter Open-Source AI Model with Self-Evolution
MiniMax M2.7 logo with a technology background

On March 18, 2026, MiniMax announced the release of M2.7, an artificial intelligence model with 230 billion parameters designed for autonomous and complex workflows. This model represents a significant advance in the field of AI, introducing self-evolution, an innovative approach that allows the model to actively take part in its own development process.

Architecture and Model Design

MiniMax M2.7 is built on a Transformer architecture with a Mixture-of-Experts (MoE) design. This approach allows the model to activate only part of its parameters during inference, improving computational efficiency. With 256 local experts and 10 billion parameters activated per token, M2.7 maintains an activation rate of 4.3%, optimizing resources during data processing.

Self-Evolution Capabilities

One of M2.7's distinctive features is its self-evolution capability. During training, the model ran more than 100 autonomous optimization cycles, analyzing its own performance, identifying areas for improvement and adapting its training structures accordingly. This approach led to a 30% performance improvement over previous versions, demonstrating the effectiveness of self-learning in continuously improving the model's capabilities.

Benchmark Performance

MiniMax M2.7 achieved notable results across various benchmarks, demonstrating its ability to tackle complex tasks in different domains:

  • SWE-Pro: 56.22%, approaching the levels of Opus 4.6.
  • Terminal Bench 2: 57.0%, demonstrating a deep understanding of complex engineering systems.
  • GDPval-AA: ELO of 1495, the highest score among open-source models in this evaluation.
  • Toolathon: 46.3%, placing it among the top performers globally.
  • Skill adherence: 97% across 40 complex skills, each with more than 2,000 tokens.

Professional Applications and Complex Environments

Beyond benchmark performance, M2.7 has demonstrated advanced capabilities in professional scenarios and complex environments. In software engineering, the model has successfully handled tasks such as end-to-end project delivery, log analysis for bug resolution, code security and machine learning. In addition, M2.7 has shown remarkable proficiency in processing professional documents, handling multi-turn edits and high-fidelity revisions in suites such as Word, Excel and PowerPoint.

Availability and Access

MiniMax M2.7 is now available on Hugging Face, with model weights made public for the research and development community. This open-source release offers opportunities for further innovation and practical applications, allowing developers to integrate M2.7 into a wide range of projects and workflows.

Conclusion

The release of MiniMax M2.7 marks a significant step in the evolution of artificial intelligence models, introducing self-evolution capabilities and advanced performance across multiple domains. With its open-source availability, M2.7 offers a powerful tool for professionals and researchers, opening new possibilities for complex and innovative AI applications.