The Sequence AI of the Week #749: Inside MiniMax-M2: Where Minimalism Meets Maximum Power
A new model that you should know about....
MiniMax-M2 stands out as one of the most technically ambitious and pragmatic open-weight models of 2025. Released by MiniMax, a Chinese AI lab with a growing reputation for engineering precision, M2 reflects a refined approach to scaling—maximizing capability for agents, coding assistants, and reasoning systems while maintaining efficient serving costs. At a time when many labs chase raw parameter counts, M2 makes the case for thoughtful architectural composition. Beneath the marketing term “Mini” lies a massive sparse Mixture-of-Experts (MoE) model, comprising around 230 billion parameters, of which only about 10 billion are active per token. This design choice forms the core of M2’s engineering philosophy: activate intelligence where needed, minimize waste elsewhere.