MiniMax M1
MiniMax · 2025-06-16 · 456.1B parameters
MiniMax M1 is MiniMax's generic hosted reasoning-model identifier and open-weight release family; its downloadable weights are the distinct 40K- and 80K-maximum-generation checkpoints, with the 40K checkpoint representing an intermediate phase of the 80K model's training. Both released checkpoints share a 32-expert MoE architecture with 456,089,655,296 parameters in the official weight files (MiniMax reports 45.9 billion active parameters per token), interleaving seven Lightning Attention blocks with one softmax-attention block and natively supporting inputs up to 1 million tokens. MiniMax continued pretraining from MiniMax-Text-01 on 7.5 trillion additional tokens, then applied supervised fine-tuning and large-scale reinforcement learning with CISPO for reasoning, software engineering, tool use, and long-context tasks.