Fish Audio releases Drama 3 weights under a revenue-capped license

On 13 October 2026, Fish Audio published Drama 3. The post says the model is available that day in Text to Speech on fish.audio, in the Fish Audio apps for iOS and Android, and through the API with the model header drama-3. It says direction is written in plain words, in English or another language, and that the model is aimed at lines that are rendered and kept: audiobooks, audio drama, game and animation dialogue, and short video. Weights are at fishaudio/drama-3.

The launch post calls the weights open weights. The license file is not an OSI-approved open-source license.

Size and architecture

The model card says Drama 3 has 35 billion parameters and 3 billion active per token, on a mixture-of-experts text backbone with hybrid full and linear attention. It describes a codebook-MTP audio decoder over 10 codebooks, FP8 expert weights, and the remaining weights in bf16. Speaker tags run from <|speaker:0|> to <|speaker:15|>, which the card says is up to 16 characters in one text.

The launch post repeats those size figures and adds a detail the card does not: the backbone was initialized from the language model of Qwen3.5-35B-A3B, with 8 of 256 routed experts plus one shared expert. config.json records model_type fish_omni, text type fish_omni_text, 40 hidden layers, 256 experts, 8 experts per token, mixed linear_attention and full_attention layers, and mixture-of-experts MLPs. The audio decoder sets decoder_type to codebook_mtp_shared and num_codebooks to 10. The file does not name Qwen and does not print a 35 billion or 3 billion total. A quantization block uses method bsgemm-moe.

The card lists 83 language codes, from zh and en through fo. The launch post says Drama 3 speaks 83 languages, and that a direction can be written in a different language from the line.

How direction is written

The launch post says direction goes in square brackets, written the way you would say it to an actor. One API example starts with [breathless, speaking fast] and later uses [voice drops to a terrified whisper]. The card’s sample uses shorter tags such as [sighs] and [laughs]. Those examples are Fish Audio’s. This story does not include a listening test.

Fish Audio says Artificial Analysis’s Pronunciation Robustness benchmark scored Drama 3 at 93.2 percent of hard words, numbers and codes, against 88.1 percent for S2.1 Pro. The Artificial Analysis page was not opened here, so those scores are Fish Audio’s account of the benchmark.

The post says that until 12 November the API is 80 percent off and new Plus, Pro or Max plans are 50 percent off, and that Drama 3 costs the same credits as S2.1 Pro. No base credit price is printed in the sections read here.

The license

The LICENSE file is the Fish Audio Community License Agreement, Version 1.0, last updated 13 October 2026. It defines Fish Audio as 39 AI, INC. and its affiliates. Research and non-commercial use are royalty-free. Commercial use is also royalty-free while annual revenue, including affiliates and from all sources, stays under 500,000 US dollars. Crossing that line requires notice to [email protected] within 30 days. The commercial grant then runs for a 90-day transition, after which commercial use needs a separate written license.

A hosted service is excluded at any revenue. The license defines that as letting third parties generate speech from text, audio or voice samples they supply, through an API, an SDK, a hosting service, or a site whose primary purpose is text-to-speech or voice cloning. Speech you generate for your own content, a studio job that delivers only finished audio, and a free non-commercial research demo are listed as outside that definition.

Section IV(b) requires prior consent before imitating an identifiable person’s voice: a parent or guardian for a minor, and the relevant rights holder for a deceased person. It also bars deceptive impersonation, sexual content involving a minor or a voice presented as a minor’s, and removing watermarks the model applies.

The card YAML says license_name: fish-audio-research-license and license_link: LICENSE.md. The file in the repo is LICENSE, and its heading is the Community License, Version 1.0. The same file says Fish Audio S2-Pro weights under the Fish Audio Research License are a different grant. The launch post says S2 stays on that research license.

Speeds Fish Audio reports

Two Fish Audio pages give different single-stream rates on one H200. The launch post says the engine decodes a single stream at 271 frames per second, against 177 for S2. The model card’s Table 1, which it says is the mean of four runs on one H200, lists 166.7 frames per second at concurrency 1, with median time to first audio of 88.9 ms. Those are Fish Audio’s measurements. A 48 GB self-host minimum was not in the card, the config or the launch post read here.