~228.7B total (MoE; 256 experts, 8 active per token; FP8 quantization)
196,608 tokens
Non-commercial MIT-style license (commercial use requires prior written authorization from MiniMax; includes attribution and use restrictions)
Open-weight MiniMax-M2.7 release positioned as a flagship agentic engineering model with strong software and productivity workflow performance. Official release materials frame M2.7 as a model that deeply participates in its own evolution, describing self-improvement loops over 100+ rounds with reported internal gains, and emphasize long-horizon task execution with Agent Teams, skill composition, and dynamic tool usage. Benchmark claims in release docs highlight strong scores across SWE-Pro, SWE Multilingual, Multi SWE Bench, VIBE-Pro, Terminal Bench 2, NL2Repo, Toolathon, GDPval-AA, and MM Claw evaluations, alongside practical production positioning for incident analysis and recovery workflows. Deployment guidance is provided for SGLang, vLLM, and Transformers, with official API and agent platform support. The later MiniMax-M2 series paper presents M2.7 as the latest checkpoint in a Mixture-of-Experts family built for agentic deployment, with the flagship M2 architecture at 229.9B total parameters and 9.8B activated per token, large-scale verifiable agent trajectories, artifact-aligned rewards, and Forge as the underlying agent-native RL system; it also specifically highlights M2.7 as an early step toward self-evolution through autonomous debugging and scaffold modification.