MiniMax opens M2.7

MiniMax has released the weights for M2.7, a software-focused artificial intelligence model, on Hugging Face as open source. (huggingface.co) Large language models predict the next token, or chunk of text, from patterns in training data. MiniMax says M2.7 is built for longer, tool-using jobs such as debugging code, editing office files, and running multi-step agent workflows. (huggingface.co, minimax.io) MiniMax and NVIDIA describe M2.7 as a sparse mixture-of-experts model, which works like a large team where only a few specialists are activated for each request. NVIDIA’s model card lists 230 billion total parameters, 10 billion active per token, 256 experts, and a context window of 204,800 tokens. (developer.nvidia.com, build.nvidia.com) MiniMax reported 56.22% on SWE-Pro and 57.0% on Terminal Bench 2, two evaluations aimed at real software engineering and terminal-based tasks rather than short code snippets. The company also reported 55.6% on VIBE-Pro and 39.8% on NL2Repo. (minimax.io, huggingface.co) Those scores put M2.7 into a crowded race around open models that try to do full engineering jobs, not just autocomplete functions. MiniMax says the model can build “agent teams,” search for tools, and carry out long chains of actions with stable role behavior. (huggingface.co) NVIDIA moved quickly to package the release for developers running graphics-processing-unit infrastructure. Its April 11, 2026 post said M2.7 is available through NVIDIA’s inference ecosystem and can be run with NemoClaw, an open source stack for managing OpenClaw assistants inside NVIDIA OpenShell. (developer.nvidia.com, docs.nvidia.com) NVIDIA’s guide also points developers to optimizations in vLLM and SGLang, two inference frameworks used to serve large models efficiently. The company said those changes target the extra routing and communication work that mixture-of-experts models require. (developer.nvidia.com) MiniMax framed the release as part of a “self-evolution” workflow, where internal versions of the model updated memory, built skills for reinforcement-learning experiments, and revised code based on evaluation results. Its Hugging Face card says one internal version improved a programming scaffold over more than 100 rounds and raised performance by 30%. (huggingface.co) The immediate test is whether outside developers can reproduce the headline benchmark numbers and turn the open weights into reliable coding products. For now, MiniMax has put M2.7 on the same public shelves where rivals are competing on cost, speed, and how much real work an open model can finish. (huggingface.co, build.nvidia.com)