MiniMax open‑sources M2.7
MiniMax published its M2.7 model as open source and says it reached state‑of‑the‑art scores on SWE‑Pro (56.22%) and Terminal Bench 2 (57.0%). The model is available now on Hugging Face for developers to inspect and use. (x.com)
MiniMax has put its M2.7 artificial intelligence model on Hugging Face and GitHub, letting developers download the weights and inspect how it was packaged. (huggingface.co, github.com) The company’s model card says M2.7 scored 56.22% on SWE-Pro, a software-engineering benchmark, and 57.0% on Terminal Bench 2, which tests work inside command-line environments. (huggingface.co, minimax.io) MiniMax says M2.7 is its first model that “deeply participat[ed] in its own evolution,” meaning the system was used during development to update memory, build tools, and refine parts of its training workflow. (huggingface.co, minimax.io) Open-weight releases let outside developers run a model on their own hardware, test its behavior, and build products without depending only on a hosted application programming interface. MiniMax’s release adds another large model to that pool at a time when many top coding systems remain closed. (huggingface.co, github.com) MiniMax is pitching M2.7 less as a chatbot than as a worker for long software and office tasks. Its model card says the system supports “Agent Teams,” dynamic tool search, and multi-step editing in Word, Excel, and PowerPoint-style workflows. (huggingface.co, minimax.io) The underlying model is large even by current standards. NVIDIA’s model card lists M2.7 as a 230 billion-parameter sparse Mixture-of-Experts system with 10 billion active parameters, a 204,800-token context window, and support for SGLang, Transformers, and vLLM. (build.nvidia.com) MiniMax’s public materials say an internal version of M2.7 optimized a programming scaffold over more than 100 rounds and improved performance by 30%, but that result comes from the company’s own reporting rather than an independent audit. (huggingface.co, minimax.io) The release is under a modified Massachusetts Institute of Technology-style license on Hugging Face, and NVIDIA’s listing says the model is ready for commercial and non-commercial use. Developers will now test whether MiniMax’s benchmark claims hold up outside the company’s own demos. (huggingface.co, build.nvidia.com)
