Fish Audio Drama 3 Is Now Open-Weight: 35B-A3B, Single-Word Edits and How to Run It

Fish Audio's Drama 3 has left preview. The model that the company announced in September behind a "comment DRAMA for a key" gate is now published as open weights on Hugging Face, with a README, serving recipes and a technical report. If you build voice apps, audiobooks, game dialogue or localization tools, this is the first open checkpoint that treats a spoken take like an editable document.

Here is the short version, then the detail: Drama 3 is a 35B-A3B mixture-of-experts text-to-speech model. It can replace one word inside a recording while the surrounding voice stays natural, revise a take from a plain instruction such as "gentler, and a bit sadder", and voice a multi-character scene from a script. We covered the September preview when almost nothing was public. This post replaces guesses with what the repository and launch coverage now say.

TL;DR: what Drama 3 is and what it costs

What changed since the September preview?

The preview was a social post and a gated API. Three things are different now. The weights are downloadable. The fish-speech repository documents installation, a command-line client, an API server and a web UI that share one interface for both Drama 3 and S2-Pro; the model type is detected from the checkpoint's config file. And there is a technical report with the evaluation details, linked from the README.

The company behind it is the same one that raised a large seed round in July. If you want the background on the business and why it keeps shipping open checkpoints, see our write-up of the Fish Audio seed round and S2.1 Pro. Note the contrast: S2.1 Pro stayed behind the paid API, while Drama 3 is the new flagship that ships with weights. AlphaSignal calls it Fish Audio's first flagship directable model in the open-weight tier, which fits the README table listing it as the "new-generation flagship".

The features, in the README's own terms

The repository lists seven key features. In plain language:

Open weights release concept showing a box opening to reveal a model checkpoint, illustrating the Fish Audio Drama 3 open weights launch

• Multi-turn director revision. You guide the previous take ("gentler, and a bit sadder") and the model regenerates it using that context rather than starting from scratch. This is the closest thing to giving notes to a voice actor.
• Infilling and precise editing. Replace a single word, the README's example is "Wednesday" to "Thursday", and the timbre and transitions around it stay natural. Co-founder and chief scientist Shijia Liao told AlphaSignal: "When you change a single word, Drama 3 listens to the audio on both sides of it and makes the new word fit."
• Scene-style multi-character production. Voice several characters from a script description, with streaming for long text.
• Long dialogues and long speech. Stable characters and timbre across extended sessions.
• Stronger expressiveness. Laughter, crying and whispers are said to be less flattened by the reference audio.
• Multilingual cloning. The README claims a new state of the art on cross-lingual error rates (character and word error rate).
• Tag control. Inline tags control emotion and prosody, and AlphaSignal describes paired tags that limit an effect to enclosed words, for example a whisper around one phrase, plus an Auto Tag helper that proposes directions for unannotated dialogue.

The word-level edit is the practical headline. Anyone who has produced audio knows the pain: a client changes one number in a script, and re-rolling the whole line changes the performance you liked. Local repair turns that into a small edit.

Drama 3 vs S2-Pro: when to pick which

The README includes its own comparison table, which we reproduce in short form:

S2-Pro is described as the previous-generation flagship with faster inference and small enough for a consumer GPU. Fish Audio reports about 100 ms to first audio and a real-time factor of 0.195 on one H200 for S2-Pro. Drama 3 trades that footprint for editing and long-form control. If you only need quick single-voice narration on a laptop-class card, S2-Pro remains the sensible pick. If you edit takes or direct scenes, Drama 3 is the point of the upgrade.

The size question: 35B total, 3B active

A mixture-of-experts router activates a subset of experts per token, so Drama 3 computes roughly like a 3B model while storing 35B parameters. That helps throughput and latency, but not memory: you still load the whole checkpoint. The same trade-off shows up in other open MoE releases we have covered, such as MiniCPM-V 4.7 35B-A3B on device and the experiment of running MoE models larger than RAM by streaming experts from SSD. For speech, where latency matters more than for chat, serving efficiency is the deciding factor, which is why the README leads with SGLang-Omni.

For a refresher on why speed numbers need context, see our explainer on tokens per second.

What is claimed versus what is verified

Some figures circulating on launch day deserve a label.

The serving numbers come from Fish Audio's own model card, so treat them as vendor claims. A fair test is simple: run the same 200-word script at one and at 64 concurrent streams on your own GPU and record time to first audio and real-time factor.

How to try it

1. Read the license. The README warns that the codebase and weights fall under the Fish Audio Community License and that S2-Pro weights stay under the Fish Audio Research License. Commercial limits are the part to check.
2. Pick a runtime. Fish Audio recommends SGLang-Omni and points to a Drama 3 README in that repository. A line from the model's repository is enough to start a server on a capable GPU box once you have installed the stack.
3. Download the checkpoint from the Hugging Face page and follow the installation docs.
4. Start with tags. Try short scripts using bracketed direction.

I'm fine. Really. <whisper>Go on without me.</whisper>
Alright, [sigh] let's try again.
[warm and relaxed, like talking to an old friend] It's been a while.

These examples come from the launch coverage, not from a test we ran. Run your own scripts before depending on them.

5. Test the edit loop. Generate a take, change one word, regenerate the region, and listen for seams. Then ask for a revision in words and see whether the second take keeps the voice.

What to test before you ship

• Consistency across runs. Does the same voice hold for ten minutes of dialogue?
• Names, numbers and dates. Those are where text-to-speech fails most, and where infilling helps most.
• Language switching. AlphaSignal says the model handles shifts across English, Chinese, Japanese and Korean within a scene.
• Cost per audio minute on your hardware versus the hosted API.
• Consent and disclosure. Voice cloning is easy to misuse; keep written permission for any voice you clone and label synthetic audio where required.

How it fits the open voice landscape

Open speech models are moving from "reads text" to "performs and edits". Fish Audio's pitch sits next to on-device recognition work such as Cactus Whistle's 16 MB speech-to-text and hosted transcription like Microsoft MAI-Transcribe-2. Together they make a complete open-ish voice loop: transcribe, edit the text, then regenerate only the changed words. The same open-release pattern appears in video, where Kandinsky 6 shipped with MIT-licensed audio-video weights.

Bottom line

Drama 3 is worth a weekend of testing if your work involves edited or directed speech. The confirmed facts are the architecture, the feature set and the public weights. The unconfirmed ones are the latency and throughput numbers and the API price. The license is the practical gate: it is open enough to download and study, but not the same as an MIT release.

Specifications reflect the Fish Audio repository and launch coverage as of October 13, 2026; model details, licensing and pricing can change.

Related reading

• Fish Audio Drama 3 preview: directing AI voice in plain language
• Fish Audio's seed round and S2.1 Pro
• Cactus Whistle: 16 MB on-device speech-to-text
• Microsoft MAI-Transcribe-2
• Kandinsky 6 open video model with audio
• MiniCPM-V 4.7 35B-A3B on device
• Sources: Fish Speech README, AlphaSignal coverage