MiniMax M2.7 Released: 229B MoE Model with Self-Evolution and Ollama Support | Model Release
MiniMax has released M2.7, a 229B parameter MoE model capable of running locally via llama.cpp and Ollama. It achieves 56.22% on SWE Pro and supports native mu…
Published on MyPrivateClaw
Apr 13, 2026, 8:37 AM UTC
Coverage date
Apr 13, 2026
Last updated
Apr 13, 2026, 8:37 AM UTC
News summary
MiniMax has released M2.7 , a 229 billion parameter mixture of experts model that runs locally via llama.cpp, Ollama, LM Studio, and SGLang. The model is available on Hugging Face and represents a significant step in open weight model capability for agentic and software engineering tasks. What Happened MiniMax published M2.7 to $1 with full weights in BF16, F8 E4M3, and F32 formats. Quantised variants for llama.cpp are already available from the community, with 37 quantisation models listed on the model page. The model supports tool calling, multi agent orchestration via native Agent Teams, and long context reasoning. M2.7 is the first model in MiniMax's lineup to incorporate what the company calls self evolution: during training, an internal version of the model autonomously optimised a programming scaffold over 100+ rounds — analysing failure trajectories, modifying code, running eval…