“The necessity of using FP32 precision during RL training was discovered during the development of Minimax M1, after a debugging process that involved a layer-by-layer analysis of log probabilities when model accuracy stalled.”