MiniMax open‑sources M2.7
MiniMax open‑sourced its M2.7 model and published results showing state‑of‑the‑art performance on SWE‑Pro (56.22%) and Terminal Bench 2 (57.0%), and made the model available on Hugging Face with API access. ( ). The release gives developers access to a high‑performing model for coding and terminal tasks outside the firm's hosted stack. (x.com)
MiniMax has released its M2.7 language model under an open-source license and posted weights on Hugging Face. (huggingface.co) The model card says M2.7 scored 56.22% on SWE-Pro and 57.0% on Terminal Bench 2, two benchmarks built around software engineering and command-line task completion. (huggingface.co, github.com) MiniMax also published the model on GitHub and says developers can use it commercially under the Massachusetts Institute of Technology license. The company’s API documentation lists M2.7 as available through its platform as well. (github.com, platform.minimax.io) A coding model is software that predicts the next token in code and text, then uses that pattern-matching to edit files, call tools, and work through bug-fixing steps. SWE-Pro measures whether a model can solve repository-level software tasks, while Terminal Bench 2 tests whether it can operate in a shell without breaking the environment. (github.com, huggingface.co) That puts M2.7 in the part of the market where developers want downloadable weights, not just a hosted chatbot. Hugging Face lists community conversions and quantized versions within hours of release, a sign that outside developers are already adapting it for local and alternative runtimes. (huggingface.co, huggingface.co, huggingface.co) MiniMax describes M2.7 as a follow-on to its earlier M2 line and says it improved “practical capabilities” over M2.5. The company’s release notes show MiniMax-M2 first launched on October 27, 2025, and the M2.7 announcement page was published on March 18, 2026. (minimax.io, platform.minimax.io, minimax.io) The published materials also pitch M2.7 for office-document editing, tool use, and multi-agent workflows, not just code generation. MiniMax says the model supports “Agent Teams” and reports additional benchmark results including 39.8% on NL2Repo and 46.3% on Toolathon. (huggingface.co, github.com) MiniMax is not the only company chasing this category, but open weights change who can experiment. With M2.7 now on Hugging Face and in MiniMax’s own API, developers can test the same model in their own stack instead of staying inside one hosted product. (huggingface.co, platform.minimax.io)
