MiniMax-M2.7 is a large language model designed for complex agentic tasks and professional productivity, featuring a novel "model self-evolution" capability. It targets engineers, researchers, and power users seeking advanced AI for software engineering, complex problem-solving, and multi-agent collaboration, offering significant performance gains through autonomous development cycles.

How It Works

M2.7's core innovation lies in its "model self-evolution" process, where the model actively participates in its own development by updating memory, constructing skills for reinforcement learning, and refining its learning based on experimental outcomes. This approach, combined with native Agent Teams, complex Skills, and dynamic tool search, enables sophisticated agent harnesses and intricate productivity task completion. An internal version demonstrated autonomous optimization, improving performance by 30% over 100+ rounds.

Quick Start & Requirements

Model weights are available on Hugging Face (MiniMaxAI/MiniMax-M2.7). Recommended inference frameworks include SGLang, vLLM, and Transformers, with dedicated deployment guides provided for each. NVIDIA NIM Endpoint is also supported. Recommended inference parameters are temperature=1.0, top_p = 0.95, top_k = 40. The default system prompt is: "You are a helpful assistant. Your name is MiniMax-M2.7 and is built by MiniMax." Further details are available via the MiniMax Agent (https://agent.minimax.io/) and API (https://platform.minimax.io/) documentation.

Highlighted Details

• Achieved a 66.6% medal rate on MLE Bench Lite, ranking second only to Opus-4.6 and GPT-5.4.
• Demonstrates strong professional software engineering capabilities, scoring 56.22% on SWE-Pro, 76.5 on SWE Multilingual, and 52.7 on Multi SWE Bench, with reported reductions in production incident recovery time to under three minutes.
• Excels in professional work tasks, achieving an ELO score of 1495 on GDPval-AA (surpassing GPT5.3), 46.3% accuracy on Toolathon, and 62.7% on the MM Claw benchmark, while handling Word, Excel, and PPT with high-fidelity editing.
• Features enhanced character consistency and emotional intelligence for entertainment applications, with an interactive demo available at openroom.ai.

Maintenance & Community

Community engagement is fostered through WeChat and Discord. For inquiries, contact model@minimax.io. Links to GitHub and ModelScope are also provided for repository access.

Licensing & Compatibility

The project includes a LICENSE file, but specific terms and compatibility for commercial use or closed-source linking are not detailed in the provided README snippet.

Limitations & Caveats

No explicit limitations, alpha status, or known bugs are mentioned in the provided documentation. Specific details regarding the license terms and potential restrictions are not elaborated upon.