NVIDIA opens MiniMax M2.7 model

NVIDIA said on April 11 that MiniMax M2.7 is now available through its platform as an open-weights model for coding and other multi-step developer tasks. (developer.nvidia.com) MiniMax M2.7 is a sparse “mixture of experts” model, which works like a team where only a few specialists answer each request instead of the whole staff. 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) NVIDIA said developers can run the model through NVIDIA Inference Microservices, or NVIDIA NIM, and pair it with NemoClaw, an open-source stack for long-running agents. NVIDIA’s NemoClaw page says the software is in early preview and installs OpenShell to add privacy and security guardrails around autonomous agents. (build.nvidia.com) (nvidia.com) The release lands as companies are trying to turn coding assistants into agents that can search logs, call tools, edit files, and keep working across longer sessions. NVIDIA’s blog says MiniMax M2.7 is aimed at software engineering, machine learning research workflows, reasoning, and office work rather than short chat prompts alone. (developer.nvidia.com) NVIDIA is not presenting MiniMax M2.7 as an in-house foundation model. Its model card says the system is “not owned or developed by NVIDIA,” and that NVIDIA is packaging third-party model access, deployment tools, and optimized inference on Blackwell and Hopper graphics processing units. (build.nvidia.com) MiniMax, the model’s developer, says M2.7 was built for “complex agent harnesses,” dynamic tool search, and “Agent Teams,” meaning multiple specialized agents working together on one job. The project’s GitHub repository says an internal version of the model improved a programming scaffold over more than 100 rounds and delivered a 30 percent performance gain in MiniMax’s internal evaluation. (github.com) MiniMax also published benchmark claims meant to place M2.7 against closed and open rivals in software work. Its GitHub page lists 56.22 percent on SWE-Pro, 57.0 percent on Terminal Bench 2, 76.5 on SWE Multilingual, and 52.7 on Multi SWE Bench, while NVIDIA’s blog says the model adds improvements over MiniMax M2.5. (github.com) (developer.nvidia.com) NVIDIA said it also worked with the open-source inference community to add performance optimizations for the MiniMax M2 series in vLLM and SGLang. The model card lists those runtimes as supported software engines on Linux systems using Hopper and Blackwell hardware. (developer.nvidia.com) (build.nvidia.com) For developers, the immediate change is not just another model download. NVIDIA is bundling model access, graphics processing unit endpoints, and an agent runtime into one stack, which makes MiniMax M2.7 a test of whether open models can handle more of the day-to-day work inside software teams. (developer.nvidia.com) (nvidia.com)