According to @_avichawla, NVIDIA Molt runs RL from a single Python module, eliminating reward models and scaling to 1T-class MoE training.