MiniMax open‑sources M2.7
MiniMax released an open‑weight M2.7 model and published benchmark results—including top scores like 56.22% on SWE‑Pro and 57.0% on Terminal Bench 2—making the model available on Hugging Face with API details for developers. The release signals another high‑performance model entering the open ecosystem and gives developers a new option for experimentation. (x.com/MiniMax_AI/status/2043132047397659000, x.com/itmedia_news/status/2043543475497623808)
MiniMax has published the weights for M2.7, putting a 230-billion-parameter model for coding and office tasks on Hugging Face and GitHub. (huggingface.co, github.com) Large language models predict the next token, or next chunk of text, from patterns learned during training; “open weights” means developers can download the model files and run them on their own systems instead of only calling a hosted application programming interface. MiniMax’s M2.7 page links both the downloadable weights and the company’s own application programming interface. (huggingface.co, github.com) MiniMax says M2.7 scored 56.22% on SWE-Pro, 57.0% on Terminal Bench 2, 55.6% on VIBE-Pro, and 39.8% on NL2Repo. The company also lists 76.5 on SWE Multilingual, 52.7 on Multi SWE Bench, and a 1495 Elo score on GDPval-AA. (huggingface.co, minimax.io) The model card describes M2.7 as a sparse Mixture-of-Experts system, a design that keeps many specialist subnetworks available but activates only a small subset for each token. NVIDIA’s model card lists about 230 billion total parameters, about 10 billion active parameters, 62 layers, and a release date of April 11, 2026. (build.nvidia.com, huggingface.co) MiniMax is pitching M2.7 as an “agent” model, meaning software that can plan steps, call tools, and work across longer tasks instead of only answering one prompt at a time. Its repository says the model supports “Agent Teams,” dynamic tool search, and more than 40 complex skills for multi-step work. (github.com, minimax.io) The company also says an internal version of M2.7 helped improve its own development setup by optimizing a programming scaffold over more than 100 rounds and lifting performance by 30%. That “self-evolution” claim comes from MiniMax’s own blog post and model card, not from an independent benchmark organizer. (minimax.io, huggingface.co) The release does not come with a standard open-source license. The Hugging Face and GitHub license files say non-commercial use is permitted on MIT-style terms, but any commercial use requires prior written authorization from MiniMax and must display “Built with MiniMax M2.7.” (huggingface.co, github.com) That puts M2.7 in the growing “open weights” category rather than the narrower open-source one recognized by the Open Source Initiative, which bars field-of-use restrictions such as a ban on commercial use. Developers can download and test M2.7 now, but companies that want to ship products on top of it need MiniMax’s approval first. (opensource.org, huggingface.co)
