Fish Audio Drama 3

Fish Audio Drama 3 is an open text-to-speech model from Fish Audio for multi-character scenes and long-form content. Per the model card, it is a mixture-of-experts model with 35 billion parameters (3 billion active per token); one text can carry up to 16 characters through speaker tags, long text is split automatically and can be streamed, and it supports voice cloning and control through natural-language instructions and tags such as [sighs] or [laughs]. The authors state a first-audio delay under 100 ms on one H200; the card does not state a VRAM requirement. The model's metadata lists 83 languages. Released on 2026-10-01; the weights are public on Hugging Face without any acceptance step. The license is the Fish Audio Community License 1.0 of 2026-10-13 (the license file in the model repository). Per the text, research and non-commercial use is free; commercial use is allowed only while annual revenue of you and your affiliates is below 500,000 US dollars, after which you must notify Fish Audio and get a separate license. A hosted service (an API, SDK or app whose main purpose is speech generation or voice cloning from other people's text and voices) needs a separate license regardless of revenue; voicing content you control and handing finished audio to a client is not a hosted service. You must display "Built with Fish Audio", may not imitate an identifiable person without consent, remove the watermark, or train other speech models on the output. You own the outputs. No territorial restrictions are in the text. The model's metadata still names an older research license, so we follow the license file. The authors offer an official demo Space on Hugging Face. The same developer's hosted service is in our catalog: Fish Audio. Date of check: 2026-10-19.