NVIDIA open-sources MiniMax M2.7

NVIDIA has added MiniMax M2.7 to its open-model lineup and is distributing it through NVIDIA services and the broader open-source inference stack. (developer.nvidia.com) MiniMax M2.7 is a text model built for long, multi-step jobs such as software engineering, tool use, reasoning, and office-document editing. NVIDIA says the release went live on April 11, 2026, and its Build platform lists a free endpoint the same day. (developer.nvidia.com, build.nvidia.com) The model uses a sparse “mixture of experts” design, which works like a large team where only a few specialists answer each request instead of the whole group. NVIDIA and MiniMax list 230 billion total parameters, 10 billion active parameters per token, 256 experts, and a context window of about 200,000 tokens. (developer.nvidia.com, build.nvidia.com) That setup is aimed at “agentic” systems, which are software agents that plan, call tools, and revise work over many steps instead of answering in one shot. NVIDIA is tying M2.7 to NemoClaw and OpenShell, its open-source and runtime tools for running always-on assistants and autonomous agents on graphics processing unit infrastructure. (developer.nvidia.com) NVIDIA is also pitching the release as an infrastructure story, not just a model launch. The company said it worked with the open-source community to add performance optimizations for MiniMax M2 models into vLLM and SGLang, two widely used inference engines. (developer.nvidia.com) MiniMax, the model developer, is presenting M2.7 as a step beyond its earlier M2.5 release. NVIDIA’s blog says M2.7 adds enhancements for reasoning, machine-learning research workflows, software engineering, and office work. (developer.nvidia.com) MiniMax’s own GitHub repository says M2.7 is its “first model deeply participating in its own evolution,” meaning the company used internal versions of the model to update memory, build skills for reinforcement-learning experiments, and refine parts of its own training process. The repository says one internal version improved a programming scaffold over more than 100 rounds and raised performance by 30%. (github.com) MiniMax also published performance claims aimed at developers choosing models for coding work. Its repository says M2.7 scored 56.22% on SWE-Pro, 52.7 on Multi SWE Bench, 57.0 on Terminal Bench 2, and 1495 Elo on GDPval-AA, which it describes as the highest among open-source models. (github.com) The Hugging Face release gives NVIDIA a familiar distribution point for developers who want to test the model outside NVIDIA’s own stack. MiniMax’s Hugging Face organization shows a MiniMax-M2.7 model updated within the last day, alongside earlier M2-series releases. (huggingface.co, huggingface.co) The immediate next step is practical adoption: developers can run M2.7 through NVIDIA endpoints, Hugging Face-hosted artifacts, or open inference frameworks that NVIDIA says it has already tuned for this model family. (build.nvidia.com, developer.nvidia.com)