MiniMax M2.7: Agent Teams and MiniMax's Self-Evolution Report

What M2.7 is

MiniMax M2.7 is an open-weight agentic model with a 200K token context window — recorded here as 204.8K, the same figure written out — priced around $0.30 per million input tokens through third-party providers. It supports native Agent Teams for multi-agent collaboration, with what MiniMax describe as stable role identity and autonomous decision-making between the members.

Self-evolution, with a number attached

MiniMax describe M2.7 as initiating a cycle of model self-evolution: during development they let the model update its own memory, build dozens of complex skills for reinforcement-learning experiments, and improve its own learning process based on the results of those experiments. An internal version of M2.7 autonomously optimised a programming scaffold across more than 100 rounds — analysing failure trajectories, modifying the code, running evaluations, and deciding on each round whether to keep or revert the change — and achieved a 30% performance improvement. It is a vendor account of an internal process with no external replication.

The engineering benchmarks

MiniMax publish a broad set of agentic and software-engineering results, all vendor-reported. On MLE Bench Lite, which spans 22 machine-learning competitions, M2.7 records a 66.6% medal rate, which MiniMax place second only to Claude Opus 4.6 and GPT-5.4. On SWE-Pro it scores 56.22%, which they describe as matching GPT-5.3-Codex. SWE Multilingual at 76.5 and Multi SWE Bench at 52.7. On VIBE-Pro it records 55.6%, which MiniMax put nearly level with Opus 4.6. Terminal Bench 2 comes in at 57.0% and NL2Repo at 39.8%.

LMArena

On LMArena's text leaderboard MiniMax M2.7 has 1415 Elo (rank 142); by category, 1480 Elo in Coding (rank 114), 1422 Elo in Math (rank 130), 1453 Elo in Expert (rank 115) and 1364 Elo in Creative Writing (rank 167).

Other figures MiniMax report

MiniMax report 46.3% on Toolathon, 97% skill compliance across more than 40 complex skills on its MM Claw benchmark, and 62.7% on MM Claw end-to-end. They describe the model handling multi-round editing of Word, Excel and PowerPoint files, log analysis, bug troubleshooting, refactoring and code security, and state that using M2.7 they reduced live production incident recovery to under three minutes on several occasions. MiniMax call M2.7 their first model deeply participating in its own evolution, and describe it as able to build complex agent harnesses using Agent Teams, complex skills and dynamic tool search, and as correlating monitoring metrics, running trace analysis and verifying root causes in databases for SRE work. MiniMax also open-sourced OpenRoom, an interactive demo that places AI interaction in a web interface with real-time visual feedback and scene interactions, which can be tried at openroom.ai.

Running it

The weights are on Hugging Face and ModelScope, and the model is offered through an NVIDIA NIM endpoint and MiniMax's own platform, including a token plan subscription. MiniMax publish deployment guides for SGLang, vLLM and Transformers and a tool-calling guide. Recommended sampling is temperature 1.0, top_p 0.95 and top_k 40, and the default system prompt identifies the model as MiniMax-M2.7 built by MiniMax.

Official Resources