Fish audio launches drama 3, a 35b moe text-to-speech model with 83 language support

Fish Audio has launched Drama 3, a 35B open-weight mixture-of-experts (MoE) text-to-speech model capable of generating acted speech from scripts with directional cues. The model supports 83 languages and allows up to 16 speakers in a single request. It is available on Hugging Face and through the Fish Audio API, with a launch price of $10 per 1M characters until November 12, 2026. Drama 3 is built on the Qwen3.5-35B-A3B language model, featuring 40 layers and a new Codebook-MTP decoder. It uses the Descript Audio Codec to output 44.1 kHz audio at 21.5 frames per second. Training involved three stages, including reinforcement learning with GRPO and a single learned human-preference reward model. The model achieved 93.2% on the Artificial Analysis Pronunciation Robustness benchmark, ranking first, and holds an Elo score of 1242 on the Artificial Analysis quality leaderboard. Fish Audio reports that the model requires at least 48 GB of NVIDIA GPU memory in FP8, with commercial use permitted under $500,000 in annual revenue. The model is licensed under the Fish Audio Community License v1.0, and users must credit the project when deploying it.