Fish Audio
company · United StatesScores
1 product on the map — 1 open-ish.
Openness
3 medium confidence- weights
- open (Drama 3 safetensors and codec on the Hub, ungated)
- data
- closed (no training-data statement for Drama 3 on its card, README or report abstract)
- code
- partial (inference code in the Fish Speech repository, served through SGLang-Omni and vLLM-Omni; no training pipeline for Drama 3)
- license
- Fish-Audio-Community-License-1.0 (Drama 3 weights and the Fish Speech code; commercial use only below USD 500,000 annual revenue, and no hosted text-to-speech service without a separate license)
- superseded-release
- S2-Pro (2026-03, 4B; weights under the non-commercial Fish Audio Research License, still published)
Anyone can download Drama 3, but its Fish Audio Community License allows commercial use only for businesses below a set annual revenue and never as a hosted voice service. Nothing is published about its training data, and only inference code is released.
- https://arxiv.org/abs/2610.18010 recorded 2026-11-12
Fish Audio Drama 3 Technical Report, submitted 13 Oct 2026: "built on a 35B-parameter mixture-of-experts backbone that activates 3B parameters per token"; "trained with a three-stage pre-training recipe, instruction fine-tuning, and large-batch reinforcement learning"; the abstract names no training corpus.
- https://huggingface.co/api/models/fishaudio/drama-3 recorded 2026-11-12
Hub record: "gated":false, codec.pth and model-00001-of-00015.safetensors through model-00015-of-00015.safetensors in the file list.
- https://huggingface.co/fishaudio/drama-3/raw/main/LICENSE recorded 2026-11-12
License file in the Drama 3 repository, titled "# FISH AUDIO COMMUNITY LICENSE AGREEMENT" / "**Version 1.0 - Last Updated: October 13, 2026**": a commercial license "but only while Your Annual Revenue is less than USD 500,000 (the "Revenue Threshold")"; "no license is granted under this Agreement to offer a Hosted Service"; "Model(s)" means the models "including Fish Audio Drama 3".
- https://huggingface.co/fishaudio/drama-3/raw/main/README.md recorded 2026-11-12
Drama 3 card: front matter "license_name: fish-audio-research-license" and "license_link: LICENSE.md" (that file returns 404), while the body reads "This model is licensed under the [Fish Audio Community License](https://huggingface.co/fishaudio/drama-3/blob/main/LICENSE)"; serving is through SGLang-Omni and vLLM-Omni.
- https://huggingface.co/fishaudio/s2-pro/raw/main/LICENSE.md recorded 2026-11-12
S2-Pro license file: "# FISH AUDIO RESEARCH LICENSE AGREEMENT" / "**Last Updated: March 7, 2026**"; "Any use of the Fish Audio Materials or Derivative Works for a Commercial Purpose requires a separate written license agreement from Fish Audio."
- https://raw.githubusercontent.com/fishaudio/fish-speech/HEAD/LICENSE recorded 2026-11-12
Code LICENSE, byte-identical to the Drama 3 license file: "Software" means Fish Audio's software "including the Fish Speech source code"; "Model weights that Fish Audio makes available under a different license (for example, the Fish Audio S2-Pro weights under the Fish Audio Research License) are not Fish Audio Materials".
- https://raw.githubusercontent.com/fishaudio/fish-speech/HEAD/README.md recorded 2026-11-12
README: "This codebase and its associated model weights are released under **[FISH AUDIO COMMUNITY LICENSE](LICENSE)**. The S2-Pro weights remain under the Fish Audio Research License."; model table "Drama 3 | 35B-A3B (3B active per token)" and "S2-Pro | 4B parameters | ... | Previous-generation flagship with faster inference"; installation and inference only.
Adoption
2 high confidenceHugging Face downloads of the Drama 3 and S2-Pro checkpoints over the trailing month. The unofficial int4 quantization and the older Fish Speech checkpoints are not counted.
- https://huggingface.co/api/models?author=fishaudio&limit=100 recorded 2026-11-12
"downloads":47450 for fishaudio/drama-3 and "downloads":46890 for fishaudio/s2-pro.
Capability
5 medium confidenceDrama 3 sits in the arena's top ten in blind listening tests, among the leading commercial voices, and is the best-rated model there whose weights can be downloaded.
- https://artificialanalysis.ai/text-to-speech/leaderboard recorded 2026-11-12
Leaderboard entry rank 7, "name":"Fish Audio Drama 3", "elo":1273.64, creator Fish Audio, "openWeightsUrl":"https://huggingface.co/fishaudio/drama-3/", "openWeights":true, "openWeightsCommercial":false; "Fish Audio S2 Pro" at rank 45; 104 models are ranked.