THE SHORT VERSION
35B total parameters, ~3B active, 83 languages, weights downloadable. First on Artificial Analysis pronunciation robustness at 93.2%. Free commercially under $500k revenue.
Fish Audio released Drama 3 on 13 October — a mixture-of-experts text-to-speech model with 35B total parameters and roughly 3B active per token, eight of 256 routed experts plus one shared. It is initialised from the language model of Qwen3.5-35B-A3B and covers 83 languages. Weights are on Hugging Face and the code on GitHub.
The benchmark claim is specific and checkable: 93.2% on Artificial Analysis' English Pronunciation Robustness benchmark, ranked first, against 88.1% for Fish's own S2.1 Pro. Inference runs at 271 frames per second on a single H200 versus 177 for S2.
The licence is the detail to read before you build on it. Fish Audio Community License v1.0 makes research and non-commercial use free, and commercial use free if you and your affiliates are under US$500,000 in annual revenue. Above that — or to offer Drama 3 as a TTS, voice-cloning or inference service — you need a licence from Fish Audio. Cloning a real person's voice without consent is banned outright. The API lists at $50 per million characters, discounted to $10 until 12 November. "Open-weight rather than open source" is Fish's own phrasing, and it is the honest one.
Sources
Fish Audio ·
MarkTechPost