MiniMax open‑sources M2.7
MiniMax released an open‑weight model called M2.7 and published the code, weights and an API on Hugging Face for developers to use and benchmark. The model hit state‑of‑the‑art scores on SWE‑Pro (56.22%) and Terminal Bench 2 (57.0%), and the release was promoted on MiniMax’s official X account. (x.com) (x.com)
MiniMax has published the weights and code for M2.7 on Hugging Face and GitHub, giving developers a new model to download, run and benchmark. (huggingface.co) (github.com) The release page lists M2.7 as a 229 billion-parameter text generation model, updated on Hugging Face within the past two days. MiniMax’s own model page says the model is also available through its application programming interface. (huggingface.co) (minimax.io) MiniMax says M2.7 scored 56.22% on SWE-Pro, 57.0% on Terminal Bench 2 and 55.6% on VIBE-Pro, all benchmarks aimed at measuring how well a model handles software engineering tasks instead of short coding puzzles. (huggingface.co) (minimax.io) Those tests matter because model makers are now competing on “agent” work: systems that use tools, inspect logs, edit codebases and complete multi-step jobs with less hand-holding than a chatbot. MiniMax says M2.7 can build “complex agent harnesses,” use dynamic tool search and coordinate “Agent Teams,” its term for multiple model roles working together. (github.com) (huggingface.co) MiniMax is also pitching M2.7 beyond coding. Its model card says M2.7 reached a 1495 Elo score on GDPval-AA for professional work tasks, 46.3% on Toolathon and 62.7% on MM Claw, while handling repeated edits in Word, Excel and PowerPoint files. (huggingface.co) (minimax.io) The company describes M2.7 as a model that took part in its own training loop. MiniMax says an internal version ran more than 100 rounds of code changes and evaluation on a programming scaffold, then kept or reverted changes, producing a 30% gain on its internal measure. (huggingface.co) (github.com) The “open-source” label comes with limits. The Hugging Face repository lists a modified MIT license, but the posted license text says commercial use requires prior written authorization from MiniMax and asks users to display “Built with MiniMax M2.7” in commercial settings. (huggingface.co 1) (huggingface.co 2) That licensing language has already become part of the story around the launch. The GitHub license file now carries the heading “NON-COMMERCIAL LICENSE,” even as the repository remains public and downloadable. (github.com 1) (github.com 2) For developers, the immediate question is less whether M2.7 is downloadable than whether its benchmark claims hold up outside MiniMax’s own materials. The weights, code and application programming interface are now public enough that outside labs can start answering that. (huggingface.co) (minimax.io)
