Fish Audio Drama 3
Fish Audio released Drama 3, an open-source speech generation model with a 35B mixture-of-experts backbone that activates 3B parameters per token and clones voices from multiple references.
Drama 3 adds multi-turn revision, scene dubbing, editing by infilling and streaming text input, all in one model. It pairs a hybrid linear-attention backbone with a new codebook multi-token-prediction audio decoder. Training uses three-stage pre-training, instruction fine-tuning and large-batch reinforcement learning with a single learned human-preference reward.
- Date
- Tuesday 13 October 2026
- Lab
- Fish Audio
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Single-stream decoding speed on one H200 | 271 frames per second Fish Audio S2 reaches 177 frames per second, with Drama 3 having more than seven times as many parameters | company |
| Parameters | 35B total, 3B active per token mixture-of-experts | company |
Weights are said to be on GitHub and Hugging Face and inference through SGLang-Omni. The license was not stated on the page read.
Sources
This record was checked against its sources on 14 October 2026. How we check