MiniMax M2.7

MiniMax M2.7 launched March 18, 2026 as the point release of the M2 series. It is a sparse mixture-of-experts model with about 230B total parameters and roughly 10B active per token, with a 200K-token context window. MiniMax says an internal version ran over 100 autonomous rounds of scaffold self-optimization, improving about 30% with no human in the loop.

The weights were released on Hugging Face under a modified MIT license whose commercial-use restriction drew community criticism; it is open-weights but not open-source. It scores 56.22% on SWE-Pro and 57.0% on Terminal Bench 2, placing it near closed frontier models on agentic coding tasks.