NVIDIA Open-Sources MiniMax M2.7
NVIDIA announced the open-source release of MiniMax M2.7, claiming state-of-the-art performance on SWE-Pro (56.22%) and Terminal Bench 2 (57.0%) with GPU-accelerated endpoints available through NemoClaw. (x.com) The release positions MiniMax as a production-capable model with specialized tooling for software-engineering benchmarks. (x.com)
NVIDIA said on April 11 that MiniMax M2.7 is now available through its stack as an open-weights model aimed at coding and long-running AI agents. (developer.nvidia.com) MiniMax M2.7 is a sparse mixture-of-experts model, which works like a large team where only a few specialists answer each prompt instead of everyone speaking at once. NVIDIA’s model card lists 230 billion total parameters, 10 billion active parameters per token, 256 experts, and a 204,800-token context window. (build.nvidia.com) The release is pitched at software work that goes beyond writing one function at a time. MiniMax’s GitHub page says the model scored 56.22% on SWE-Pro and 57.0% on Terminal Bench 2, two benchmarks built around debugging, repository work, and command-line tasks. (github.com) NVIDIA said MiniMax M2.7 is available across the open-source inference ecosystem and highlighted support work in vLLM and SGLang, two serving frameworks used to run large models on graphics processing units. The company also said developers can run the model through NemoClaw, an open-source reference stack for always-on assistants built on OpenShell. (developer.nvidia.com) MiniMax describes M2.7 as a model that helped improve its own training workflow. On its GitHub and Hugging Face pages, the company says an internal version optimized a programming scaffold over more than 100 rounds and improved performance by 30%. (github.com, huggingface.co) The model is also being marketed as a work tool, not only a coding tool. MiniMax says M2.7 reached 1495 Elo on GDPval-AA, 46.3% on Toolathon, and 62.7% on its MM Claw end-to-end benchmark, alongside support for multi-agent “Agent Teams.” (huggingface.co) The licensing picture is less settled than the launch language suggests. Hugging Face labels the repository “modified-mit,” while the current GitHub LICENSE file says commercial use requires prior written authorization and permits only non-commercial use on MIT-style terms. (huggingface.co, github.com) That leaves developers with two separate questions after the benchmark claims: how well M2.7 holds up outside vendor-run tests, and whether its latest license terms fit production use. NVIDIA’s own model card says the model is “ready for commercial/non-commercial use,” but also notes that the model is owned by a third party and governed by separate license terms. (build.nvidia.com)
