Fish Audio Drama 3 is an open-source speech generation model built on a 35B-parameter mixture-of-experts backbone with 3B active parameters per token. It supports multi-reference voice cloning, natural-language performance direction, multi-turn revision, multi-character scenes, speech infilling, and long-form or streaming generation. A hybrid linear-attention backbone and codebook multi-token-prediction decoder enable efficient inference, while a three-stage pre-training pipeline, instruction fine-tuning, and large-batch reinforcement learning align the model with human preferences.