MiniMax M2

MiniMax M2, published on 23 October 2025, is the point at which the lab stopped scaling its flagship up. Its predecessor carried 456 billion parameters with 45.9 billion active per token; M2 carries 230 billion with 10 billion active — half the model, and a fraction of the compute per token. MiniMax describes it as “a Mini model built for Max coding and agentic workflows”, and the design follows: 62 layers, 256 experts with eight routed per token, a routing pattern far more granular than the two-of-32 used in the previous generation. The trade was context. Text-01 and M1 were built around a one-million-token window; M2 ships with 196,608 positions, roughly a fifth of that, and the hybrid lightning attention that made the long window affordable is gone from the configuration. MiniMax would not return to a million tokens until M3 in May 2026, and only then with a new sparse-attention design. In exchange the model became cheap to run and easy to deploy, and it anchored a line rather than a single release: M2.1 in December 2025, M2.5 in February 2026 and M2.7 in April 2026 all reuse the same 230-billion-parameter body, with M2.7 nudging the window to 204,800. The weights are published under a modified MIT licence — more permissive than the custom agreement on Text-01, less than the plain Apache 2.0 of M1. Independent providers list M2 at about $0.26 per million input tokens and $1.02 per million output.

▸Videos

No videos yet.

▸Reviews