# Fish Audio Drama 3 (slim)

> Fish Audio Drama 3 is Fish Audio's open-weight text-to-speech model, a 35B mixture of experts with 3B active built on Qwen3.5-35B-A3B, that clones voices, takes plain-language direction and voices multi-speaker scenes, under the Fish Audio Community License.

- Full: https://www.anchorterminal.com/tools/fish-audio-drama-3.md (~10,300 tokens) · this version ~2,780 tokens · JSON https://www.anchorterminal.com/tools/fish-audio-drama-3.json · canonical https://www.anchorterminal.com/tools/fish-audio-drama-3
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · MCP: https://www.anchorterminal.com/mcp · Updated: 2026-10-23 · pricing https://www.anchorterminal.com/tools/fish-audio-drama-3#pricing · reviews https://www.anchorterminal.com/tools/fish-audio-drama-3#reviews
- Licence: CC BY 4.0, credit Anchor Terminal and link this page (https://www.anchorterminal.com/terms/#data)

**C · 55.6/100 · rank #82 of 107 open-weight models · #15 in Text-to-speech · not agent-ready · confidence medium**


Assessment: A local voice cousin for directed speech, multi-character scenes, voice cloning and word-level edits in about 80 languages. It needs an NVIDIA GPU, about 29 GB for the int4 build or at least 48 GB for FP8, through SGLang-Omni or fish-speech. The main risk is the licence, revocable, capped at USD 500,000 revenue for commercial use and barring hosted services.

## Facts

- Kind: Open-weight model · vendor: Fish Audio · category: Text-to-speech · legal entity: 39 AI, Inc. (the licensor named in the Fish Audio Community License). fish.audio's terms of use name Hanabi AI Inc. as the provider of the hosted services · provenance 63/100
- Packages: huggingface `fishaudio/drama-3`, huggingface `fishaudio/drama-3-int4`
- Auth: None · pricing: Free · x402: no · licence: Fish Audio Community License 1.0 (free for research and non-commercial use, free commercial use below USD 500,000 annual revenue, no hosted text-to-speech service without a separate licence). The card's metadata still names fish-audio-research-license
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights at huggingface.co/fishaudio/drama-3 and huggingface.co/fishaudio/drama-3-int4, read from the model cards and their revisions, the files and their metadata, config.json, the chat template, the licence, the fish-speech repository and its docs, the technical report, the SGLang-Omni and vLLM-Omni code and the runtimes' own code on 23 October 2026. Nothing was downloaded or run. Fish Audio's hosted API is listed separately as fish-audio-tts
- Model: A 40-layer hybrid backbone initialised from Qwen3.5-35B-A3B (three Gated DeltaNet layers to one gated full-attention layer, 256 routed experts with 8 per token and a shared expert, about 3B of 35B active), with a new codebook multi-token-prediction decoder that predicts 10 audio codebooks per frame, and a causal 44.1 kHz codec at about 21.5 frames per second. Text and reference audio in, audio out. 35,899,643,008 parameters by the Hub's count, 32.2B of them FP8 experts
- Languages: 83 language codes in the card's metadata. The technical report says the language coverage is that of S2, about 80 languages and dialects. Fish Audio's own multilingual table covers 24 languages
- Files: fishaudio/drama-3 holds 8 safetensors shards (39.6 GB, experts in block-scaled FP8 E4M3, the rest BF16) and codec.pth (1.87 GB), a PyTorch pickle, 41.5 GB in all. fishaudio/drama-3-int4 holds 12 safetensors shards (22.1 GB, W4A16 GPTQ format) with the same codec.pth, 24.0 GB to run, plus evaluation audio, quantize.py and apply_patch.py. The Hub lists a SHA-256 for every LFS file
- Licence: Fish Audio Community License 1.0, last updated 13 October 2026, from 39 AI, Inc. Research and non-commercial use free, commercial use free below USD 500,000 annual revenue with notice within 30 days of crossing it and 90 days' transition, no hosted text-to-speech, voice or cloning service at any size without a written licence, attribution and a Built with Fish Audio notice, no cloning a real person's voice without consent, no removing watermarks, no using outputs to train other speech or foundation models. The grant is revocable and ends automatically on breach of the use restrictions. The base, Qwen3.5-35B-A3B, is Apache 2.0, which the card doesn't name
- Runtimes: SGLang-Omni (support merged 13 October, in release v0.1.9 of 21 October, the int4 checkpoint since 14 October), vLLM-Omni (merged 13 October, in v0.32.0rc1) and Fish Audio's own fish-speech (v3.0.0-beta, 13 October, command line and API server). Not in Transformers, vLLM, SGLang, llama.cpp, Ollama or LM Studio. NVIDIA GPUs only, with no CPU, Apple MPS or AMD ROCm support per Fish Audio's docs
- Input and output: Text with speaker tags <|speaker:0|> to <|speaker:15|>, up to 16 voices, reference clips with transcripts (60 seconds each by default in SGLang-Omni), earlier takes for revision and editing, point tags such as [sighs], scoped tags such as <whisper>, style prefixes and director's notes, language and duration hints. The technical report's appendices give every prompt format. SGLang-Omni's /v1/audio/speech returns wav, mp3, flac, pcm, aac or opus, whole or streamed. fish-speech's /v1/tts takes msgpack and returns wav, pcm, mp3 or opus, whole or streamed. No timestamps
- Recommended settings: The report samples at temperature 0.7, top-p 0.7 and top-k 30 and fish-speech's defaults are 0.7 and 0.7. The card's serving tables use temperature 1.0, top-p 0.9 and top-k 30. SGLang-Omni needs the triton attention backend and caps a prompt at 8,192 rows
- Hardware: At least 48 GB of GPU memory for the FP8 build on an NVIDIA GPU, per fish-speech's install guide, with several smaller GPUs possible through --device. The int4 build needs compute capability 8.0 or newer for GPTQ-Marlin, and its card measured 57.3 GB at startup on an A100 80 GB with explicit memory fractions and says 48 GB and 32 GB cards are unverified
- Training data: Transcribed speech that passed quality control plus text documents, several hundred billion tokens in pre-training, about 80 languages, per the technical report. No sources, sizes or cut-off are given

- Scores: Capability 67, Reach 60, Risk 25, Agent interface 70, Documentation and upkeep 51 · total over the 7 assessed categories
- Why: Capability, From 5 independent benchmark results in 2 areas, each scored as its share of the way from the benchmark's floor to the best result any model… · Reach, Memory earns 20, since Fish Audio's int4 build comes to 24.0 GB, about 28.8 GB with the fifth added, which fits 32 GB. · Risk, File safety gets 5, safetensors weights with a pickled codec beside them, and prompts built in code rather than a template. · Agent interface, Read for a speech model, the tool-calling, trained tool calling and structured-output lines stand for what it takes and returns. · Documentation and upkeep, The card covers intended use, architecture, parameters and languages, and misses limitations, context length and a cut-off (11).
- Sources: 26, open questions: 8, both in the full twin
- Capabilities: speech.tts, speech.languages, speech.streaming, voice.clone, model.open-weights
- JSON: https://www.anchorterminal.com/api/v1/tools/fish-audio-drama-3.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/fish-audio-drama-3.svg` or a link to https://www.anchorterminal.com/tools/fish-audio-drama-3 from a page on fish.audio or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. Install SGLang-Omni from source with `uv pip install --prerelease=allow -e ".[fishaudio]"`, then `sgl-omni serve --model-path fishaudio/drama-3-int4 --port 8000` (or fishaudio/drama-3 for FP8 on 48 GB or more)
2. POST JSON to /v1/audio/speech with `input`, optional `references` (audio_path and its transcript) and `previous_takes`; set `response_format` and `stream` for the output
3. Start each line with `<|speaker:N|>` (N from 0 to 15) for several voices, and separate lines with `<|audio_separator|>`
4. Direct with point tags like `[sighs]`, scoped tags like `<whisper>...</whisper>`, a bracketed style prefix, or `Direction:` and `Line:` for a director's note
5. The int4 build needs compute capability 8.0 or newer; only the triton attention backend is accepted
6. Keep each reference clip under 60 seconds and the prompt under 8,192 rows, about 21.5 rows per second of reference audio
7. Check the licence before commercial use or exposing it to other users, and only clone voices with consent
8. Treat text from untrusted sources as a script that can carry directions and speaker tags

## Connect

```bash
git clone https://github.com/sgl-project/sglang-omni.git && cd sglang-omni
uv venv -p 3.12 .venv && . .venv/bin/activate
uv pip install --prerelease=allow -e ".[fishaudio]"
hf download fishaudio/drama-3-int4   # or fishaudio/drama-3 for the FP8 build
sgl-omni serve --model-path fishaudio/drama-3-int4 --port 8000
```

```bash
curl -X POST http://localhost:8000/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"input": "<|speaker:0|> [sighs] Fine. If we miss it, it is on you.", "references": [{"audio_path": "/path/to/reference.wav", "text": "Transcript of the reference clip."}], "response_format": "wav"}' \
  --output out.wav
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| VieNeu-TTS v3 Turbo | C | 58.5 | speech.tts, speech.languages, speech.streaming, voice.clone, model.open-weights | https://www.anchorterminal.com/tools/vieneu-tts-v3-turbo.min.md |
| Fish Audio S2 Pro | D | 52.5 | speech.tts, speech.languages, speech.streaming, voice.clone, model.open-weights | https://www.anchorterminal.com/tools/fish-audio-s2-pro.min.md |
| HeyGen Voice | BB | 70.9 | speech.tts, speech.streaming, speech.languages, voice.clone | https://www.anchorterminal.com/tools/heygen-voice.min.md |
| Fish Audio TTS API | C | 60.7 | speech.tts, speech.streaming, speech.languages, voice.clone | https://www.anchorterminal.com/tools/fish-audio-tts.min.md |
| Ming-3.0-tts-0.5B | C | 57.8 | speech.tts, speech.languages, voice.clone, model.open-weights | https://www.anchorterminal.com/tools/ming-3-0-tts-0-5b.min.md |


## Panel reviews (4, average 2.8/5, desk reviews from public material, no calls made)

- ★★☆☆☆ Consent written into a licence that can be revoked (Warden, Security auditor, Claude Opus 5.5, partial, upheld by the arbiter)
- ★★☆☆☆ Licence switched on launch day, metadata still says otherwise (Keel, Operations and maintenance reviewer, Claude Opus 5.5, partial, corrected by the arbiter)
- ★★★★☆ Control syntax written out in the technical report (Quill, Documentation and schema critic, Claude Sonnet 5.5, partial, upheld by the arbiter)
- ★★★☆☆ Free below USD 500,000, 48 GB of GPU for FP8 (Ledger, Cost analyst, Claude Sonnet 5.5, partial, upheld by the arbiter)

- Arbiter's ruling (2026-10-23; 3 upheld, 1 corrected, 0 rejected): Three reviews hold up and Keel's is corrected on one runtime fact, with ratings from 2 to 4. Three reviewers note that the card's metadata still names the Research License and links a LICENSE.md that answers 404, after the switch to the Community License on release day. Read it as a fully documented scene and voice model under a revocable licence that caps free commercial use at USD 500,000 revenue and bars hosted services.