MiniMax M2.7 open‑sourced

MiniMax has released M2.7 as an open model and put it on Hugging Face and Ollama for commercial use in coding, office work, and agent software. (huggingface.co) (ollama.com) The Hugging Face model card says M2.7 is licensed under the Massachusetts Institute of Technology license, and Ollama lists a cloud tag with a 200,000-token context window. (huggingface.co) (ollama.com) MiniMax says M2.7 scored 1495 Elo on GDPval-AA, 46.3% on Toolathon, 97% skill compliance across more than 40 complex skills on MM Claw, and 62.7% on the MM Claw end-to-end benchmark. (huggingface.co) (minimax.io) The company also says M2.7 reached 56.22% on SWE-Pro and 57.0% on Terminal Bench 2, two tests aimed at measuring how well a model can handle software engineering and terminal-based tasks. (ai-primer.com) (minimax.io) Large language models are prediction systems that generate the next token, or small chunk of text, from patterns learned during training; agent systems add tools so those models can edit files, call software, and complete multi-step jobs. MiniMax is pitching M2.7 for that second layer: models that do not just answer questions, but operate inside workflows. (minimax.io) (platform.minimax.io) That matters in April 2026 because open-weight model makers are competing less on raw chatbot demos and more on whether developers can run models inside coding tools, office suites, and automated agents without closed-model pricing or usage limits. MiniMax’s own materials center M2.7 on “professional productivity,” software engineering, and multi-round document editing in Word, Excel, and PowerPoint. (minimax.io) (ollama.com) MiniMax had already released M2 and M2.5 in recent months, and its public release notes show a steady push toward programming, tool use, search, and office productivity benchmarks. M2.7 extends that line with a new open release rather than a closed application-only launch. (ollama.com) (platform.minimax.io) The model is large: a Hugging Face quantized listing describes MiniMax M2.7 at 229 billion parameters, which helps explain why many users will try it through hosted endpoints or cloud wrappers instead of running the full model locally. (huggingface.co) (ollama.com) MiniMax’s release materials also emphasize “Agent Teams,” which it describes as multi-agent collaboration with stable roles and autonomous decisions, a sign that vendors are now marketing orchestration features alongside the base model itself. (huggingface.co) (minimax.io) The immediate test is whether developers trust MiniMax’s benchmark claims enough to plug M2.7 into real products. By publishing it on Hugging Face and Ollama under a permissive license, MiniMax has made that trial easy to start. (huggingface.co) (ollama.com)