MiniMax M2.7 Is Now Open Source
7 hours ago
- #AI Self-Evolution
- #Software Engineering AI
- #Agentic Models
- MiniMax allowed an AI model, M2.7, to self-evolve by modifying its own code through unsupervised analysis of failures, resulting in a 30% performance improvement without human intervention.
- M2.7 achieved a 66.6% average medal rate on MLE Bench Lite over three 24-hour trials, showing continuous improvement, second only to models like Opus 4.6 and GPT-5.4.
- In software engineering benchmarks, M2.7 matches GPT-5.3-Codex on SWE-Pro and performs well on multilingual and real-world scenarios, with claims of reducing incident recovery times to under three minutes.
- For productivity, M2.7 excels in document editing for Word, Excel, and PPT, scoring high on GDPval-AA and tool use accuracy, making it suitable for office tasks.
- The license allows free non-commercial use but requires prior written authorization from MiniMax for commercial use, including display of "Built with MiniMax M2.7" attribution.
- M2.7 is available via NVIDIA's free API, HuggingFace for local deployment with recommended frameworks like SGLang, and hosted interfaces for testing agentic capabilities.