NVIDIA open-sources MiniMax M2.7
NVIDIA announced MiniMax M2.7 as an open-source model tuned for agentic workflows and complex reasoning, with state-of-the-art results on several coding and terminal benchmarks. The release comes with guides for running the model on NVIDIA endpoints and tooling like NemoClaw and OpenClaw. (x.com, developer.nvidia.com)
An artificial intelligence “agent” is a language model wired to use tools, keep track of steps, and finish multi-part jobs like writing code or operating a terminal. NVIDIA said on April 11 that MiniMax M2.7 is now available as open weights on its platform and through public repositories. (developer.nvidia.com, build.nvidia.com) MiniMax M2.7 is a sparse “mixture of experts” model, which means it has many specialist sub-models but only activates a few for each token to cut compute. NVIDIA’s model card lists 230 billion total parameters, 10 billion active parameters per token, 256 experts, and a 200,000-token context window. (build.nvidia.com) NVIDIA’s April 11 post said developers can run the model through NVIDIA Inference Microservices, or NVIDIA NIM, and can fine-tune it with the open-source NVIDIA NeMo AutoModel library. The same post said NVIDIA added performance work for the MiniMax M2 series to vLLM and SGLang, two popular open-source inference engines. (developer.nvidia.com) The pitch is not a general chatbot but a model for long-running software tasks. NVIDIA said M2.7 posted state-of-the-art results on Terminal-Bench and SWE-bench Verified among open models, benchmarks that test whether a model can complete command-line and software-engineering work. (developer.nvidia.com) MiniMax’s Hugging Face page frames the same model around office and agent work, saying it can edit Word, Excel, and PowerPoint files across multiple rounds and scored 46.3% on Toolathon and 62.7% on MM Claw. Those tests are meant to measure whether a model can pick tools, follow long task plans, and keep skills consistent over many steps. (huggingface.co) The release also gives NVIDIA a new outside model to showcase its own agent stack. Its blog points developers to NemoClaw, an open-source reference stack, and OpenClaw, an always-on assistant setup that runs inside NVIDIA OpenShell, which the company describes as a secure runtime for autonomous agents. (developer.nvidia.com, developer.nvidia.com) MiniMax, the company behind the model, published the weights on Hugging Face and code on GitHub, while NVIDIA published deployment guides and hosted endpoints. That split lets NVIDIA sell infrastructure and tooling around a model it did not train itself. (huggingface.co, github.com, developer.nvidia.com) The repositories describe M2.7 as MiniMax’s first model to “participate” in its own development by updating memory, building skills for reinforcement-learning experiments, and improving parts of its training loop. That claim is part of the model’s branding, and NVIDIA repeated it in its launch materials without publishing a separate audit of the process. (github.com, developer.nvidia.com) For developers, the immediate change is practical: a large agent-focused model is now downloadable, benchmarked, and packaged with NVIDIA’s serving and orchestration tools as of April 11. For NVIDIA, it is another way to tie open models to its chips, runtimes, and agent software. (build.nvidia.com, developer.nvidia.com)
