MiniMax reports that the M1 model was trained using large-scale reinforcement learning (RL) at an efficiency rarely seen in this domain, with a total cost of $534,700.