MiniMax-M2.7 is built on a Sparse Mixture-of-Experts architecture that activates roughly 10 billion parameters per token while maintaining access to approximately 230 billion total parameters—this mixture design lets the model keep inference costs dramatically lower than dense models of comparable capability. The model is designed first and foremost for agentic workflows: it natively supports tool calling and MCP, enabling it to bolt on image understanding or web search through external tools, and it orchestrates Agent Teams with complex Skills and dynamic tool search to complete elaborate productivity tasks. MiniMax positions it as built for "top real-world engineering" and "professional office delivery," targeting the kind of multi-step coding and productivity work that requires a model to manage scaffolding, evaluate its own outputs, and iterate without constant human guidance.

What sets M2.7 apart from earlier MiniMax models is its deep participation in its own evolution—during development, an internal version autonomously ran over 100 optimization cycles on a programming scaffold, analyzing failure trajectories, modifying code, running evaluations, and deciding whether to keep or revert changes, which yielded a 30% performance improvement without human intervention. The model also independently managed a significant portion of its own reinforcement learning research workflows and competed in 22 ML competitions on MLE-Bench Lite, winning 9 gold medals and achieving a 66.6% medal rate that places it second only to Opus-4.6 and GPT-5.4 on that benchmark. In real-world head-to-head testing against Claude Opus 4.6, M2.7 delivered approximately 90% of the quality at about 7% of the total task cost while running three times faster, making it particularly attractive for teams that need frontier-adjacent coding performance without frontier-level pricing.