NVIDIA open-sources MiniMax M2.7
NVIDIA released MiniMax M2.7 as an open model and made GPU‑accelerated endpoints available for experimentation, aiming to support agentic workflows and software engineering tasks. The company highlighted benchmark results—SWE‑Pro 56.22% and Terminal Bench 2 at 57.0%—and pointed users to GPU‑accelerated endpoints via tools like NemoClaw. (x.com)
NVIDIA has put MiniMax M2.7 on its platform as an open model and exposed a free GPU-backed endpoint for testing on NVIDIA NIM. (developer.nvidia.com) (build.nvidia.com) MiniMax M2.7 is a text model built for coding, reasoning, and office tasks; NVIDIA’s April 11, 2026 post lists it at 230 billion parameters, with 10 billion active per token, 256 experts, and a 200,000-token context window. (developer.nvidia.com) NVIDIA said developers can run the model through NVIDIA NIM endpoints or wire it into NemoClaw, an open-source stack that installs OpenShell, a sandboxed runtime for autonomous software agents. (developer.nvidia.com) An “agent” here means a model that can call tools, keep state across steps, and act inside a controlled environment instead of stopping at a single text reply. NVIDIA’s pitch is that M2.7 fits those longer software and productivity workflows. (developer.nvidia.com) The benchmark numbers NVIDIA highlighted come from MiniMax’s own model page: 56.22% on SWE-Pro, 57.0% on Terminal Bench 2, and 55.6% on VIBE-Pro. MiniMax describes those tests as measures of software engineering, terminal-based system work, and end-to-end project delivery. (minimax.io) (developer.nvidia.com) The release also extends a recent pattern on NVIDIA’s model hub, where the company hosts third-party open models behind managed inference endpoints while steering developers toward its own deployment stack. NVIDIA’s model catalog now lists MiniMax M2.7 as a free endpoint and MiniMax 2.5 as a downloadable model. (build.nvidia.com 1) (build.nvidia.com 2) MiniMax, the model maker, separately published the weights on Hugging Face under a modified MIT-style license, where the repository says M2.7 “deeply participat[ed] in its own evolution.” That language points to MiniMax’s effort to train models that help build and evaluate later versions of themselves. (huggingface.co) (unite.ai) License terms are part of the story. NVIDIA’s endpoint page says use of the hosted trial is governed by NVIDIA API Trial Terms, while the model itself is governed by the NVIDIA Open Model License Agreement; Decrypt reported MiniMax also changed some commercial terms shortly after release. (build.nvidia.com) (decrypt.co) For developers, the immediate change is practical: M2.7 can now be tried on NVIDIA-hosted GPUs before anyone decides whether to download weights, quantize them, or fit the model into a larger agent system. (build.nvidia.com) (developer.nvidia.com)
