MiniMax M2.7 open‑sourced

A large language model is a system that predicts the next token, or chunk of text, and developers often release the model weights so others can run it themselves. MiniMax has now published M2.7 on Hugging Face and GitHub with downloadable weights and deployment guides for vLLM, SGLang, and Transformers. (huggingface.co) (github.com) M2.7 is a sparse mixture-of-experts model, which works like a panel of specialists where only a few are consulted for each token instead of the whole network firing every time. NVIDIA says the model has 230 billion total parameters, 10 billion active parameters per token, 256 experts, and a 200,000-token context window. (developer.nvidia.com) MiniMax says M2.7 scored 56.22% on SWE-Pro and 57.0% on Terminal Bench 2, two tests aimed at software engineering and command-line task completion. The company also reported 76.5 on SWE Multilingual, 52.7 on Multi SWE Bench, and 55.6% on VIBE-Pro. (huggingface.co) (minimax.io) Those numbers target a crowded part of the market: open models that can do more than chat and can instead write code, call tools, inspect logs, and finish multi-step jobs. MiniMax says M2.7 was built for “agent harnesses,” meaning software setups where the model can use tools and coordinate longer workflows. (huggingface.co) (github.com) MiniMax also tied the release to office-style work, not just coding. The model card says M2.7 reached an Elo score of 1495 on GDPval-AA, handled Word, Excel, and PowerPoint editing, and scored 46.3% on Toolathon with 97% skill compliance across more than 40 complex skills on MM Claw. (huggingface.co) (minimax.io) The company framed the model as part of a “self-evolution” loop during training and evaluation. The Hugging Face card says an internal version of M2.7 optimized a programming scaffold over more than 100 rounds and improved performance by 30%. (huggingface.co) NVIDIA used the launch to promote the infrastructure around these models as much as the model itself. Its April 11, 2026 post said M2.7 is available through NVIDIA and the open-source inference ecosystem, and pointed developers to NemoClaw, an open-source stack for running always-on assistants with the OpenShell runtime. (developer.nvidia.com) NVIDIA also said it worked with the open-source community on vLLM and SGLang optimizations for MiniMax’s architecture, including kernels for Query-Key Root Mean Square Normalization and floating-point 8 mixture-of-experts inference. That puts M2.7 in the same race as other open models competing on both benchmark scores and serving efficiency. (developer.nvidia.com) The release is open-source in the “weights are public” sense, but the license is not unrestricted. The Hugging Face page lists a modified-MIT license, and the repository license says non-commercial use is permitted while commercial use requires prior written authorization from MiniMax. (huggingface.co) (github.com) So the immediate change is concrete: developers can download M2.7 now, test MiniMax’s benchmark claims, and decide whether its coding-and-agent focus is worth building around. The harder question, which the next few weeks of third-party evaluations will answer, is whether those published scores hold up outside MiniMax’s own reports. (huggingface.co) (developer.nvidia.com)