NVIDIA open‑sources MiniMax M2.7

NVIDIA has added MiniMax M2.7 to its open model lineup, publishing the weights on April 11 and offering a free hosted endpoint to test it. (developer.nvidia.com) (build.nvidia.com) MiniMax M2.7 is a text model built for coding, tool use, and office work. NVIDIA’s model card lists it at 230 billion total parameters, with 10 billion active per token, a 204,800-token context window, and support for SGLang, Transformers, and vLLM on Hopper and Blackwell graphics processors. (build.nvidia.com) A “mixture of experts” model works like a team where only a few specialists answer each request instead of waking up the whole staff. NVIDIA said MiniMax M2.7 uses 256 experts and activates eight per token, which is how it keeps inference costs lower than a dense model of the same total size. (developer.nvidia.com) (build.nvidia.com) The release is aimed at developers building software agents, which are systems that can search, run tools, edit files, and keep working across many steps. NVIDIA said the model is tuned for long-horizon software engineering, live production troubleshooting, and document generation and editing. (build.nvidia.com) (developer.nvidia.com) MiniMax’s own model card pitches M2.7 as stronger on real engineering work than plain code completion. It reports 56.22% on SWE-Pro, 57.0% on Terminal Bench 2, 39.8% on NL2Repo, and 76.5 on SWE Multilingual, alongside claims of sub-three-minute incident recovery in some live production cases. (huggingface.co) (github.com) Those benchmark numbers come from the model publisher, not from an independent audit in NVIDIA’s announcement. NVIDIA’s post says the company also worked with the open-source serving community to add performance optimizations for MiniMax M2 models in vLLM and SGLang, including kernels for query-key normalization and floating-point 8 mixture-of-experts inference. (developer.nvidia.com) The licensing is more open than MiniMax’s earlier API-only positioning but not a public-domain release. NVIDIA’s model card says the hosted trial is covered by NVIDIA API Trial Terms, while use of the model itself is governed by the NVIDIA Open Model License and a modified Massachusetts Institute of Technology license. (build.nvidia.com) The practical effect is that developers can now pull the model from Hugging Face, run it through open inference stacks, or try it through NVIDIA’s endpoint without setting up a full deployment first. That turns MiniMax M2.7 from a benchmark claim into something engineers can test against their own codebases now. (huggingface.co) (build.nvidia.com)