otheraudioRuns locally35B MoE
Drama 3: Open TTS Specs, License & How to Run
Open-weight 35B MoE text-to-speech model (~3B active) with a directed-performance decoder.
- License
- other
- Parameters
- 35B MoE (active ~3B)
- HF likes
- 112 (2026-10-14)
- First seen
- 2026-10-01
TL;DR. Fish Audio Drama 3 is a audio model from Fish Audio (39 AI, INC.). Parameters: 35B MoE (active ~3B).License: other. It is listed under ElevenLabs.
| Developer | Fish Audio (39 AI, INC.) |
|---|---|
| Type | audio |
| Architecture | Hybrid linear-attention MoE backbone initialized from Qwen3.5-35B-A3B, plus a codebook-MTP audio decoder; ModifiedDAC codec, 10 codebooks, ~21.5 Hz, 44.1 kHz output |
| Parameters | 35B MoE · active ~3B |
| Modalities | audio, speech, text |
| Languages | 83 languages (Fish Audio blog) |
| License | other — Fish Audio Community License Agreement Version 1.0 (last updated 13 October 2026). Free for research and non-commercial use. Free commercial use only while Annual Revenue of You and Affiliates is under USD 500,000. Offering a Hosted Service (TTS/voice API to third parties) needs a separate Fish Audio license regardless of revenue. Cloning a real person’s voice without consent is banned. Contact [email protected]. The Hugging Face cardData still labels license_name fish-audio-research-license / LICENSE.md; the file on the repo is Community License v1.0. Not legal advice. |
| Hardware | 48 GB (FP8, official docs) — Official install.md (checked 2026-10-14) requires an NVIDIA GPU and at least 48 GB VRAM for Drama 3 FP8. It states no CPU, Apple MPS, or AMD ROCm for Drama 3, and no --compile. forkalt has not measured VRAM. (not independently tested) |
| Hugging Face likes | 112 (as of 2026-10-14) |
| Model card | https://huggingface.co/fishaudio/drama-3 |
| First seen | 2026-10-01 |
Fish Audio Drama 3 (drama-3) is an open-weight text-to-speech model. Official hosts we opened on 2026-10-14: the ungated card fishaudio/drama-3, the inference repository fishaudio/fish-speech, GitHub prerelease v3.0.0-beta, and the launch post Fish Audio Drama 3 (13 October 2026).
This identification page hangs off open ElevenLabs alternatives. That mapping is editorial. We have no sourced quality comparison against ElevenLabs.
The blog describes directed performance (plain-language tags in square brackets), 83 languages, and open weights under a new Community License. The GitHub prerelease calls it “an open-source speech generation model that turns text into directed performances.” Hugging Face pipeline_tag is text-to-speech. Those are copied labels, not a ranking.
What we verified (2026-10-14)
| Field | Value | Source |
|---|---|---|
| Hugging Face likes | 112 | HF API fishaudio/drama-3 |
| HF downloads (monthly counter) | 23 | HF API |
| Card created | 2026-10-01 | HF createdAt |
| Card last modified | 2026-10-13T16:06:45Z | HF lastModified |
| Gated | false | HF API |
| License file | Fish Audio Community License Agreement Version 1.0, last updated 13 October 2026 | LICENSE |
| HF cardData license_name | fish-audio-research-license / LICENSE.md |
HF cardData — stale vs the LICENSE file and blog |
| GitHub stars / forks | 33,061 / 2,871 | GitHub API fishaudio/fish-speech |
| Release | v3.0.0-beta, “Fish Audio Drama 3 — Pre-Release”, prerelease, 2026-10-13T16:05:05Z |
GitHub Releases |
| Params | 35B MoE, ~3B active per token (8 of 256 routed experts + shared); initialized from Qwen3.5-35B-A3B | Blog + GitHub release notes |
| Codec | ModifiedDAC, 10 codebooks, ~21.5 Hz, 44.1 kHz output | Blog / release |
| Languages | Blog: 83; HF card lists many ISO language codes | Blog + HF cardData.language |
| Minimum GPU (official) | NVIDIA only; ≥48 GB VRAM for FP8; no CPU/MPS/ROCm | install.md |
| Download | hf download fishaudio/drama-3 --local-dir checkpoints/drama-3 |
same install.md |
Architecture (from the blog and release notes)
The Fish Audio blog and the v3.0.0-beta notes describe a 35B-parameter mixture-of-experts backbone with about 3B active per token (8 of 256 routed experts plus a shared expert), initialized from the language model of Qwen3.5-35B-A3B. Attention is hybrid: Gated DeltaNet linear-attention layers with one gated full-attention layer in four. A codebook-MTP audio decoder replaces S2’s per-frame Fast AR. The tokenizer is described as S2’s, built on the Descript Audio Codec: 44.1 kHz, ten codebooks, about 21.5 frames per second.
We copy those figures. We have not counted parameters from the safetensors shards.
How to run (from official install.md)
docs/en/install.md in fishaudio/fish-speech (checked 2026-10-14) documents Linux/WSL, apt install portaudio19-dev libsox-dev ffmpeg, then conda or uv with a CUDA extra (cu126 / cu128 / cu129). Drama 3 is NVIDIA GPU only. The documented download is:
hf download fishaudio/drama-3 --local-dir checkpoints/drama-3If codec.pth is missing from that directory, the same page says the S2-Pro codec at checkpoints/s2-pro/codec.pth is used. Docker images fishaudio/fish-speech:server-cuda are documented. forkalt has not run these commands.
The blog also documents a hosted API (POST https://api.fish.audio/v1/tts with header model: drama-3). That is Fish Audio’s product, not the local checkpoint.
License
The file at huggingface.co/fishaudio/drama-3/resolve/main/LICENSE is titled Fish Audio Community License Agreement Version 1.0, last updated October 13, 2026. Attribution in that file names 39 AI, INC.
Hugging Face cardData still says license_name: fish-audio-research-license and license_link: LICENSE.md. The blob on the repo is LICENSE, and the blog says Community License v1.0. We follow the file and the blog, and we record the stale cardData label so you do not mix it with S2’s Research License.
Section III (paraphrase, not legal advice): commercial use is licensed only while Annual Revenue of You and Affiliates is under USD 500,000. Hosted Service (making the materials available so third parties can generate speech from inputs they supply) needs a separate Fish Audio license regardless of revenue. The license bans cloning a real person’s voice without consent. Contact [email protected].
What we did not verify
Artificial Analysis Pronunciation Robustness 93.2% overall (blog table; also cited as #1 open-weights TTS / Pronunciation Robustness, with an Artificial Analysis post) is vendor-cited / third-party. forkalt has not re-run the bench. Inference speed 271 frames/s on one H200 is from the blog and release notes, not our stopwatch. We have no sourced bake-off against ElevenLabs.
Related speech on this site: pyannote community-1 via Otter.ai alternatives (diarization, not TTS). Video generation is on Runway alternatives.
Benchmarks
Quoted from the cited source. forkalt has not independently reproduced these numbers.
| Name | Score | Source |
|---|---|---|
| Artificial Analysis Pronunciation Robustness (overall, Drama 3 Preview) | 93.2% (self-reported) | https://fish.audio/blog/fish-audio-drama-3/ |
Deploy options
Comparisons
FAQ
Are Drama 3 weights actually public?
Yes. Hugging Face fishaudio/drama-3 is ungated (created 2026-10-01; 112 likes and 23 downloads in the 2026-10-14 API check). Code is github.com/fishaudio/fish-speech. GitHub published prerelease v3.0.0-beta “Fish Audio Drama 3 — Pre-Release” at 2026-10-13T16:05:05Z. The Fish Audio blog dated 13 October 2026 says the model is available on fish.audio, the iOS/Android app, and the API.
Can I use Drama 3 commercially?
The LICENSE file on the Hugging Face repo is titled Fish Audio Community License Agreement Version 1.0 (last updated 13 October 2026). Section III licenses commercial use only while Annual Revenue of You and Affiliates is under USD 500,000. Hosted Services (TTS/voice APIs offered to third parties) need a separate Fish Audio license regardless of revenue. Cloning a real person’s voice without consent is banned. Contact [email protected]. This is not legal advice.
Does Drama 3 beat ElevenLabs?
We do not claim that. This page identifies an open-weight TTS checkpoint. The Fish Audio blog cites Artificial Analysis Pronunciation Robustness at 93.2% overall for Drama 3 Preview and a ranking as #1 open-weights TTS / Pronunciation Robustness. Those are vendor-cited third-party numbers. forkalt has not re-measured them and has no sourced bake-off against ElevenLabs.
What hardware does the official install guide require?
docs/en/install.md in fishaudio/fish-speech (checked 2026-10-14) says Drama 3 needs an NVIDIA GPU and at least 48 GB VRAM for FP8, with no CPU, Apple MPS, or AMD ROCm. Download is `hf download fishaudio/drama-3 --local-dir checkpoints/drama-3`. We have not run that command.
Sources & verification
- https://huggingface.co/fishaudio/drama-3
- https://github.com/fishaudio/fish-speech
- https://github.com/fishaudio/fish-speech/releases/tag/v3.0.0-beta
- https://huggingface.co/fishaudio/drama-3/resolve/main/LICENSE
- https://fish.audio/blog/fish-audio-drama-3/
- https://raw.githubusercontent.com/fishaudio/fish-speech/main/docs/en/install.md
- https://x.com/ArtificialAnlys/status/2110081938140582085
License notes are not legal advice. Read the original LICENSE. Trademark & disclaimer.