MiniMax open‑sources M2.7

MiniMax has published M2.7 for download on Hugging Face and GitHub, putting one of its latest coding-focused models into public hands. (huggingface.co) (github.com) The company says M2.7 scored 56.22% on SWE-Bench Pro, a test that asks models to fix real GitHub software issues, and 57.0% on Terminal-Bench 2, which measures how well an agent works inside a command-line environment. (minimax.io) (scaleapi.github.io) (tbench.ai) (swebench.com) MiniMax first announced M2.7 on March 18, 2026 in its model release notes and now links the release across its API docs, model card, and code repository. The Hugging Face page lists the model as 229 billion parameters with custom code support, while third-party deployment guides describe it as 230 billion total parameters with 10 billion active at a time. (platform.minimax.io) (huggingface.co) (unsloth.ai) These benchmarks matter because they test a newer style of artificial intelligence system: not just a chatbot that writes text, but an “agent” that reads files, runs commands, edits code, and checks whether its own work passed. SWE-Bench is built from real software bugs, and Terminal-Bench uses sandboxed terminal tasks such as system administration and model training. (swebench.com) (scaleapi.github.io) (vals.ai) MiniMax describes M2.7 as a model that took part in its own training workflow by updating memory, building skills for reinforcement-learning experiments, and revising parts of its harness after results came back. In its model card, the company says an internal version optimized a programming scaffold over more than 100 rounds and improved performance by 30%. (minimax.io) (huggingface.co) The release also comes with deployment and tool-calling guides, which signals that MiniMax wants developers to run the model rather than just read about it. The GitHub repository includes documentation for Transformers, vLLM, and SGLang, alongside a separate tool-calling guide. (github.com 1) (github.com 2) MiniMax says M2.7 also targets office software and multi-agent work, with a 1495 Elo score on GDPval-AA, 46.3% on Toolathon, and 62.7% on MM Claw. On the same model card, the company says M2.7 supports “Agent Teams,” or multiple model roles working together on one task. (huggingface.co) (github.com) The release is not fully open-source in the standard software sense. The license posted on Hugging Face and GitHub permits non-commercial use, requires attribution for commercial use, and says any commercial use or commercial derivative work needs prior written authorization from MiniMax. (huggingface.co) (github.com) That leaves M2.7 in a familiar 2026 category: downloadable weights, public benchmarks, and runnable code, but with licensing terms that still keep commercial control with the model maker. For researchers and hobbyists, the immediate next step is simple: download it, test the benchmark claims, and see how far the agent setup actually goes. (huggingface.co) (github.com)