MiniMax

by MiniMax · Website

MiniMax's M-series are open-weight frontier Mixture-of-Experts models built around MiniMax Sparse Attention for efficient long context. MiniMax M3 (June 2026) pairs 428B total / 23B active parameters with a 1M-token window and native image and video understanding — the first open-weight release to combine frontier coding with true multimodality. Weights ship under the MiniMax Community License (permissive with attribution; large companies need authorization).