MiniMax M2.7 open‑sourced

MiniMax has released M2.7 as an open model, putting a 230 billion-parameter system for coding and tool use on Hugging Face and GitHub. (huggingface.co) (github.com) MiniMax says M2.7 is a sparse “mixture of experts” model, which works like a team where only a few specialists answer each token: 230 billion total parameters, 10 billion active per token, 256 experts, and a 200,000-token context window. (developer.nvidia.com) The company published benchmark scores aimed at software and agent work: 56.22% on SWE-Pro, 57.0% on Terminal Bench 2, 55.6% on VIBE-Pro, 46.3% on Toolathon, and 62.7% on MM Claw. (minimax.io) (huggingface.co) Those tests measure how well a model handles real programming tickets, command-line jobs, and multi-step tool use rather than short chat prompts. SWE-bench describes its core task as fixing real GitHub issues in code repositories, and Terminal Bench 2 says its tasks run inside containerized environments such as debugging code and resolving security flaws. (github.com 1) (github.com 2) MiniMax is pitching M2.7 less as a chatbot than as a worker model for “agentic” software, meaning systems that can call tools, edit files, and keep going across long tasks. Its model card also says the release targets office software workflows including Word, Excel, and PowerPoint editing. (huggingface.co) NVIDIA moved quickly to tie the launch to its own stack. In a blog post published two days ago, NVIDIA said M2.7 can run on NVIDIA Inference Microservices and with NemoClaw, an open-source reference stack for OpenClaw assistants built on the NVIDIA OpenShell runtime. (developer.nvidia.com) (github.com) NemoClaw’s GitHub page says the software is in early preview starting March 16, 2026, and is designed to run always-on assistants “more safely” inside NVIDIA’s environment. That gives MiniMax a distribution path beyond model downloads and into managed endpoints and enterprise deployments. (github.com) (build.nvidia.com) The release also extends MiniMax’s recent push into open-weight models for coding. In November 2025, the company released MiniMax-M2 with the same 230 billion total and 10 billion active-parameter design, calling it a model for coding and agentic workflows. (github.com) What happens next is less about one benchmark table than whether developers actually build on it. MiniMax has now put M2.7 in the two places open-model users look first — Hugging Face for weights and GitHub for code — while NVIDIA is offering the infrastructure pitch alongside it. (huggingface.co) (github.com) (developer.nvidia.com)