MiniMax open‑sources M2.7
MiniMax has open‑sourced a new M2.7 model that the briefings say achieves top scores on developer benchmarks like SWE‑Pro and Terminal Bench 2, and Nvidia highlighted GPU‑accelerated endpoints and tools such as NemoClaw for agentic workflows. The release is positioned for developers experimenting with complex AI apps and shows continued momentum in community‑facing model releases. (x.com/NVIDIAAIDev/status/2043133171949264901)
MiniMax has released M2.7 as an open model aimed at coding agents and complex workplace tasks. (minimax.io) MiniMax said M2.7 scored 56.22% on SWE-Pro, 57.0% on Terminal Bench 2, and 55.6% on VIBE-Pro, three tests used to measure software engineering and terminal-based task performance. (minimax.io) The company’s model page describes M2.7 as a mixture-of-experts system with 230 billion total parameters, 10 billion active parameters per token, 256 experts, and a 200,000-token context window. (developer.nvidia.com) A mixture-of-experts model works like a large team where only a few specialists handle each step, which cuts the computing cost of running a very large system. MiniMax said that setup is meant for “agent teams,” tool use, and long jobs that move across code, documents, and terminals. (github.com) MiniMax said M2.7 was its first model to “participate in its own evolution” by updating memory, building skills for reinforcement-learning experiments, and improving parts of its own training loop from experiment results. (github.com) The release lands as more model makers are publishing weights and code for developers who want to run systems outside a closed application programming interface. MiniMax posted M2.7 on both GitHub and Hugging Face, alongside a technical report and model materials. (github.com, huggingface.co) Nvidia tied the model to its own developer stack in a blog post published April 11, 2026, highlighting hosted Nvidia Inference Microservices endpoints and NemoClaw, an open-source setup for running always-on assistants with the OpenShell runtime. (developer.nvidia.com, developer.nvidia.com) MiniMax also pushed M2.7 beyond coding. The company said the model reached a 1495 Elo score on GDPval-AA, which it described as the highest among open-source models for office-style work such as Word, Excel, and PowerPoint editing. (minimax.io, huggingface.co) Independent caution still applies to all vendor benchmark claims. Nvidia’s model card says MiniMax M2.7 is a third-party model “not owned or developed by NVIDIA,” even as Nvidia offers it through its own deployment tooling. (build.nvidia.com) For developers, the immediate question is less whether M2.7 wins every leaderboard than whether an openly released 230 billion-parameter agent model is now practical to test in real workflows. MiniMax and Nvidia are both betting that answer is yes. (minimax.io, developer.nvidia.com)
