MiniMaxAI's MiniMax-M2.7 is a 228.7 billion parameter language model designed for self-evolution and complex agentic tasks, featuring a 32768 token context length. It excels in professional software engineering, achieving high performance on benchmarks like SWE-Pro and VIBE-Pro, and supports multi-agent collaboration. The model is optimized for productivity tasks, dynamic tool use, and demonstrates strong reasoning capabilities in real-world scenarios.
Loading preview...
MiniMax-M2.7: A Self-Evolving Agentic Language Model
MiniMax-M2.7 is a 228.7 billion parameter model from MiniMaxAI, distinguished by its self-evolution capabilities. During its development, the model updated its own memory, built complex skills for reinforcement learning experiments, and improved its learning process based on results. An internal version autonomously optimized a programming scaffold over 100+ rounds, achieving a 30% performance improvement.
Key Capabilities & Performance
- Self-Evolution: Demonstrated a 30% performance improvement in programming scaffold optimization through autonomous analysis and iteration.
- Agentic AI: Capable of building complex agent harnesses, leveraging Agent Teams for multi-agent collaboration, and dynamic tool search.
- Professional Software Engineering: Excels in log analysis, bug troubleshooting, refactoring, code security, and machine learning. Achieved 56.22% on SWE-Pro (matching GPT-5.3-Codex), 76.5 on SWE Multilingual, and 52.7 on Multi SWE Bench.
- System-Level Reasoning: Strong in correlating monitoring metrics, trace analysis, and root cause verification, reducing live production incident recovery time to under three minutes on multiple occasions.
- Productivity Tasks: Achieved an ELO score of 1495 on GDPval-AA (highest among open-weight models) and handles high-fidelity multi-round editing for Word, Excel, and PPT.
- Tool Use: Reached 46.3% accuracy on Toolathon and maintains 97% skill compliance across 40+ complex skills on MM Claw.
- Entertainment: Features strengthened character consistency and emotional intelligence, showcased in the interactive demo OpenRoom.
When to Use This Model
- Complex Software Development: For advanced code generation, debugging, refactoring, and system-level engineering tasks.
- Agentic Workflows: Ideal for building multi-agent systems, automating complex productivity tasks, and dynamic tool integration.
- High-Fidelity Document Editing: For professional-grade editing and generation of Word, Excel, and PowerPoint documents.
- Research in Self-Evolving AI: Offers insights into models capable of improving their own learning and performance.