MiniMax open‑sources M2.7

MiniMax has open-sourced M2.7, a large language model aimed at coding and long-running software agents, with weights posted on Hugging Face on April 11. (huggingface.co; docs.api.nvidia.com) A large language model predicts the next token in a sequence; an “agent” wraps that model with tools like a shell, browser, or code runner so it can take multi-step actions. MiniMax says M2.7 is built for those tool-using workflows, including software debugging, document editing, and production troubleshooting. (huggingface.co; docs.api.nvidia.com) The model uses a sparse mixture-of-experts design, which works like a large team where only a few specialists answer each request. Nvidia says M2.7 has 230 billion total parameters, activates 10 billion per token, uses 256 experts with 8 active at a time, and supports a 200,000-token context window. (developer.nvidia.com; docs.api.nvidia.com) MiniMax reported 56.22% on SWE-Pro and 57.0% on Terminal Bench 2, two benchmarks that test whether a model can fix real software problems and operate in a terminal instead of just autocomplete code. The Hugging Face model page also lists 76.5 on SWE Multilingual, 52.7 on Multi SWE Bench, and 39.8 on NL2Repo. (huggingface.co) Those scores matter because open-weight coding models have increasingly been judged on end-to-end tasks, not just code snippets. SWE-Pro and Terminal Bench 2 measure whether a model can inspect files, run commands, recover from errors, and finish a job inside a realistic engineering loop. (huggingface.co) MiniMax also framed M2.7 as a model for “Agent Teams,” its term for multiple model roles working together with stable identities and tool access. On the same model card, the company said an internal version of M2.7 improved a programming scaffold over more than 100 rounds and raised performance by 30%. (huggingface.co) Nvidia moved quickly to package the release for developers running experiments on its hardware. Its April 11 technical post said M2.7 is available through Nvidia services and described optimizations in vLLM, SGLang, and TensorRT-LLM for mixture-of-experts inference. (developer.nvidia.com; docs.api.nvidia.com) Nvidia’s API documentation says the hosted model is ready for commercial and non-commercial use, while also noting that MiniMax owns the model and that use is governed by MiniMax’s license terms alongside Nvidia’s trial terms. The same page lists support for Linux and Nvidia Blackwell and Hopper graphics processing units, including B100, B200, GB200, H100, and H200. (docs.api.nvidia.com) The release adds another open-weight option in a market where developers increasingly want downloadable models they can benchmark, fine-tune, or run behind their own firewalls. For now, the headline is simple: MiniMax is trying to compete in coding agents with an open model, and Nvidia is helping put it on graphics processing units fast. (huggingface.co; developer.nvidia.com)