The paper shows how to speed up reinforcement learning (RL) for large language models (LLMs) by making numbers smaller (FP8) without breaking training.
HeartMuLa is a family of open-source music AI models that can understand and generate full songs with clear lyrics and strong musical structure.