Drama 3
Fish Audio's 36B mixture-of-experts text-to-speech model for directed performances. 83 languages, free below US$500K revenue.
In short. Drama 3 turns a script into acted speech. You write directions such as [whispering] or [laughs] in the text, and it can voice several characters in one scene and copy voices from short clips. It is a large model, and the license is free for commercial use only for businesses under US$500,000 in yearly revenue.
Summary
Drama 3 is Fish Audio's text-to-speech model for directed performances, released on 13 October 2026. Fish Audio is a voice AI company that also sells the model through its own app and API. The weights are on Hugging Face under the Fish Audio Community License Agreement 1.0, which is free for research and non-commercial use, and free for commercial use only while you and your affiliates earn less than US$500,000 a year. Offering Drama 3 to others as a text-to-speech or voice-cloning service is not licensed at any size. The model speaks 83 languages, clones voices from one or more reference clips, and takes direction in square brackets such as [sighs] or [whispering], which can be written in a different language from the text.
It is a mixture-of-experts model with 35.9 billion parameters and about 3 billion active per token, using 8 of 256 routed experts plus one shared expert. Its backbone was initialized from the language model of Qwen3.5-35B-A3B and has 40 layers in blocks of three Gated DeltaNet layers and one gated full-attention layer. A new codebook multi-token-prediction decoder predicts the ten audio codebooks (44.1 kHz, about 21.5 frames per second) and attends over the whole sequence, including the reference clips. One text can carry up to 16 speaker tags for multi-character scenes, long text is split into chunks automatically and can be streamed, and an earlier take can be sent back as context to revise it. Training had three pre-training stages, instruction fine-tuning and reinforcement learning with GRPO and a single learned human-preference reward. The expert weights are FP8, so the main checkpoint is 41.5 GB.
Fish Audio says Drama 3 reads 93.2% of hard words correctly on Artificial Analysis' Pronunciation Robustness benchmark, the top score on that leaderboard, against 88.1% for its earlier S2.1 Pro; we have not checked the leaderboard. On one H200 GPU the card reports a first audio after 89 ms for a single stream and a real-time factor of 0.80 at 64 concurrent streams. SGLang-Omni and vLLM-Omni serve it. A second repository in Fish Audio's organization holds a 4-bit build of 22.1 GB that its card calls an unofficial quantization. It runs on A100-class GPUs, and in the quantizer's own test on one H200 it had a word error rate of 1.63% against 1.99% for the FP8 build. The Hugging Face metadata names the license "fish-audio-research-license", but the LICENSE file is the Community License, and the older S2 weights stay under the Research License.
Good for
- Audiobooks, audio drama, games and dubbing where each line needs its own emotional direction
- Multi-character scenes of up to 16 voices from one script
- Self-hosted voice generation for your own content or product while you stay under US$500,000 in yearly revenue
- Research on large text-to-speech models, since the technical report and weights are public
Watch out for
- Commercial use above US$500,000 in yearly revenue needs a separate license, and a hosted text-to-speech or voice-cloning service needs one at any size
- The license bans cloning a real person's voice without consent, removing watermarks, and training other speech or foundation models on its outputs
- It is heavy, with 41.5 GB of weights, so by our estimate you need a GPU with about 48 GB for the FP8 build, or about 25 GB for the unofficial INT4 build
- The pronunciation ranking comes from Fish Audio's own post, and the editing tools shown at launch were still in invite-only testing
- Hugging Face metadata and the LICENSE file name the license differently, so read the file itself
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of Drama 3. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| Drama 3 FP8 | 35.9B 3B active | 42 GB | 48 GB | Workstation |
| Drama 3 INT4 (hybrid W4A16, unofficial quantization) INT4 | 35.9B 3B active | 22 GB | 25 GB | Gaming GPU |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
Free for research and non-commercial use, and free for commercial use while you and your affiliates earn less than US$500,000 a year. Above that, or to offer the model to others as a text-to-speech, voice-cloning or inference service at any size, you need a separate written license from Fish Audio. You must show "Built with Fish Audio", you may not clone a real person's voice without consent, and you may not use the model or its audio to train other speech or foundation models.
- Attribution or model name must be shown
- Free commercial use only below a revenue cap
- Separate deal needed to sell hosted access above a threshold
- Acceptable-use policy applies
More about Fish Audio Community 1.0
See also
- ZONOS3-Mini vs Drama 3: which new open TTS model?
- Best open-source speech models for transcription and text-to-speech
- Best open-source text-to-speech models in 2026
- Free until you grow: where open licenses draw the line
- Open speech models: transcription is open, text-to-speech is mixed
- ZONOS3-Mini: open TTS in 43 languages that runs in a browser
- Open licenses: free to build on, but hosting needs a deal
Frequently asked questions
- What is Drama 3?
- Drama 3 turns a script into acted speech. You write directions such as [whispering] or [laughs] in the text, and it can voice several characters in one scene and copy voices from short clips. It is a large model, and the license is free for commercial use only for businesses under US$500,000 in yearly revenue.
- Can I use Drama 3 commercially?
- Yes, with conditions. Drama 3 is released under the Fish Audio Community 1.0, which allows commercial use but adds extra rules: Attribution or model name must be shown; Free commercial use only below a revenue cap; Separate deal needed to sell hosted access above a threshold; Acceptable-use policy applies. Read the license text before you ship a product.
- Is Drama 3 open source?
- Drama 3 is open-weight, not open source in the strict sense. You can download the model files, but the Fish Audio Community 1.0 adds conditions that the Open Source Initiative definition does not allow.
- Can I run Drama 3 on my own computer?
- Yes, on a good graphics card. The smallest variant, Drama 3 INT4 (hybrid W4A16, unofficial quantization), needs about 25 GB of memory at 4-bit precision, which suits a 24 to 32 GB gaming GPU or a Mac with 32 GB or more.
- When was Drama 3 released?
- Drama 3 was released by Fish Audio on October 13, 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Fish Audio Drama 3: The Most Emotive Text-to-Speech Model We've Built (Fish Audio blog, 13 Oct 2026) fish.audio
- Drama 3 model card (Hugging Face) huggingface.co
- Fish Audio Drama 3 Technical Report (arXiv 2610.18010) arxiv.org
- Fish Audio Community License Agreement 1.0 (Hugging Face) huggingface.co
- Drama 3 INT4 model card (Hugging Face) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| S2 Pro Fish Audio | 4.56B | – | Fish Audio Research (Non-commercial) | 24 GB Gaming GPU | |
| ZONOS3-Mini Zyphra | 800M 100M active | – | Apache 2.0 (Permissive) | 10 GB Laptop or phone | |
| VibeVoice-TTS Microsoft | 1.02B – 2.7B | 64K | MIT (Permissive) | 2.35 GB Laptop or phone | |
| Qwen3-TTS Alibaba Qwen | 910M – 1.92B | – | Apache 2.0 (Permissive) | 3.68 GB Laptop or phone |