MiniMax Open‑Sources M2.7

MiniMax has released the weights for M2.7, an open model it says tops current open-source results on software and agent benchmarks. (minimax.io) The company reports 56.22% on SWE-Pro, a test of fixing real software bugs, and 57.0% on Terminal Bench 2, which measures how well a model works through command-line engineering tasks. MiniMax posted the model on Hugging Face and GitHub, and NVIDIA listed April 11, 2026 as the Hugging Face release date on its model card. (minimax.io) (huggingface.co) (build.nvidia.com) These tests matter because they score systems on multi-step coding work, not just short answers. MiniMax says M2.7 also reached 55.6% on VIBE-Pro, 39.8% on NL2Repo, 76.5 on SWE Multilingual, and 52.7 on Multi SWE Bench. (huggingface.co) (github.com) M2.7 is a sparse mixture-of-experts model, which works like a large team where only a few specialists handle each request. NVIDIA says the model has 230 billion total parameters, activates 10 billion per token, uses 256 experts with 8 active per token, and supports a 200,000-token context window. (developer.nvidia.com) (build.nvidia.com) MiniMax is pitching that design for “agentic” work, meaning software that can plan, use tools, and carry out longer jobs with limited human input. Its model card says M2.7 supports “Agent Teams,” dynamic tool search, and complex skills for coding, office documents, and production troubleshooting. (huggingface.co) (github.com) The company also tied the release to a broader deployment push. NVIDIA said developers can run M2.7 through NVIDIA Inference Microservices, or NVIDIA NIM, and use NVIDIA NemoClaw with OpenClaw and OpenShell to launch always-on assistants on Brev cloud graphics processing unit infrastructure. (developer.nvidia.com) (build.nvidia.com) NVIDIA said it worked with the open-source serving community to add performance optimizations for vLLM and SGLang, two widely used inference frameworks. Those changes target the model’s mixture-of-experts routing and normalization steps, which are the parts that decide which specialists wake up for each token. (developer.nvidia.com) MiniMax says M2.7 was also used during its own development. The company wrote that an internal version optimized a programming scaffold over more than 100 rounds and improved performance by 30%, though that claim comes from MiniMax’s own model card rather than an independent benchmark. (huggingface.co) (github.com) The release lands as model makers compete on open weights, coding benchmarks, and turnkey deployment instead of chat quality alone. MiniMax’s bet is that an open model with strong tool use, long context, and day-one infrastructure support can win developers who want agents they can run themselves. (huggingface.co) (developer.nvidia.com)