MiniMax open‑sources M2.7
MiniMax AI released its M2.7 model as open source and reports strong benchmark results on developer and coding tests. The company cited scores like 56.22% on SWE‑Pro and 57.0% on Terminal Bench 2 as evidence of the model's competitiveness for infrastructure and consulting use cases (x.com).
MiniMax has published the weights for M2.7, a new coding-focused language model, on GitHub and Hugging Face after first introducing the system on March 18. (github.com) (minimax.io) Large language models predict the next token in a sequence, and coding models are tuned to keep that prediction chain coherent across files, terminals, and debugging steps. MiniMax says M2.7 is built for that longer-horizon work, including bug hunting, log analysis, code security, and machine learning tasks. (minimax.io) (github.com) MiniMax reported 56.22% on SWE-Pro, 55.6% on VIBE-Pro, and 57.0% on Terminal Bench 2, three tests aimed at software engineering and command-line problem solving. The company also said M2.7 reached 76.5 on SWE Multilingual and 52.7 on Multi SWE Bench. (github.com) (minimax.io) The model uses a mixture-of-experts design, which works like routing a task to a few specialists instead of waking up the whole workforce every time. MiniMax and model listings describe M2.7 as a roughly 229 billion-parameter system built for agent workflows rather than simple chat replies. (unite.ai) (together.ai) MiniMax’s main pitch is not just benchmark scores but “self-evolution,” its term for letting an internal version of the model update memory, build skills, and revise parts of its own training workflow. The company said one internal M2.7 variant ran more than 100 optimization rounds on a programming scaffold and improved internal programming performance by 30%. (minimax.io) (github.com) MiniMax also tied M2.7 to office and agent software, not only software engineering. It said the model scored 1495 Elo on GDPval-AA, handled Word, Excel, and PowerPoint editing, and maintained 97% skill adherence across more than 40 complex skills. (minimax.io) (github.com) The release lands in a market where Chinese labs have been publishing more downloadable model weights to win developers and enterprise users. MiniMax said M2.7 had day-one support across serving stacks including SGLang, vLLM, Ollama Cloud, Together AI, and NVIDIA NIM. (ai-primer.com) (huggingface.co) But the release is not open source in the standard software sense if commercial use needs separate approval. The license posted with the model says any commercial use or derivative commercial use requires prior written authorization from MiniMax and asks users to display “Built with MiniMax M2.7.” (huggingface.co) (github.com) That licensing wrinkle means researchers can inspect and run the weights, while companies that want to ship products on top of M2.7 still need MiniMax’s permission. For now, the clearest fact is that MiniMax has put a high-scoring coding model into public repositories while keeping the commercial gate in its own hands. (huggingface.co 1) (huggingface.co 2)
