Fish Audio has released Drama 3, a 35 billion parameter open-weight mixture-of-experts text-to-speech model designed to interpret plain-language acting directions.
Key points
- Fish Audio released Drama 3, a 35B open-weight MoE text-to-speech model.
- Drama 3 interprets acting directions provided in bracketed cues.
- The model supports 83 languages and up to 16 speakers.
- It requires 48 GB of VRAM on an NVIDIA GPU.
Drama 3 accepts scripts with bracketed cues, such as [barely holding it together], to guide the speech delivery. Paired tags like <whisper> can modify specific words, and directions can be in any language, independent of the spoken line.
The model supports 83 languages and allows for up to 16 speakers per request using speaker tags <|speaker:0|> to <|speaker:15|>. The API processes a full scene in a single call by assigning one voice ID per speaker.
By the numbers
- Total parameters
- 35B
- Active parameters per token
- 3B
- Minimum GPU VRAM required
- 48 GB
Built on the Qwen3.5-35B-A3B language model, Drama 3 features 40 layers with a novel Codebook-MTP decoder. It requires at least 48 GB of VRAM on an NVIDIA GPU for operation.
While Fish Audio's highest-scoring model on Artificial Analysis benchmarks, Drama 3's quality Elo is behind several closed APIs. It achieved 95.9% on brands, places, and technical terms but scored 86.3% on exact sequences like codes and paths.
Sources and references
Original report
Fish Audio Releases Drama 3, a 35B Open-Weight MoE Text-to-Speech Model
MarkTechPost
Read on MarkTechPost (opens in a new tab)Linked in the original report
- Linkbeta.fish.audio/blog/fish-audio-drama-3 (opens in a new tab)
- Modelhuggingface.co/fishaudio/drama-3 (opens in a new tab)
- Linkdocs.fish.audio/features/text-to-speech/drama-3 (opens in a new tab)
- Linkfish.audio/blog/fish-audio-open-sources-s2 (opens in a new tab)
- Linkartificialanalysis.ai/text-to-speech (opens in a new tab)
- Linkartificialanalysis.ai/text-to-speech/leaderboard/provider-voice (opens in a new tab)
Ref 00TNO07S
Cite this
APA
Interestana AI Editorial. (2026, October 14). Fish Audio Releases Drama 3 Text-to-Speech Model. Interestana. https://interestana.com/articles/fish-audio-releases-drama-3-a-35b-open-weight-moe-text-to-speech-model-00tno07s
MLA
“Fish Audio Releases Drama 3 Text-to-Speech Model.” Interestana, 14 Oct. 2026, interestana.com/articles/fish-audio-releases-drama-3-a-35b-open-weight-moe-text-to-speech-model-00tno07s.
Chicago
Interestana AI Editorial. “Fish Audio Releases Drama 3 Text-to-Speech Model.” Interestana, October 14, 2026. https://interestana.com/articles/fish-audio-releases-drama-3-a-35b-open-weight-moe-text-to-speech-model-00tno07s.
BibTeX
@misc{interestana_00tno07s,
author = {{Interestana AI Editorial}},
title = {{Fish Audio Releases Drama 3 Text-to-Speech Model}},
howpublished = {Interestana},
year = {2026},
month = oct,
url = {https://interestana.com/articles/fish-audio-releases-drama-3-a-35b-open-weight-moe-text-to-speech-model-00tno07s},
note = {Summary of reporting by MarkTechPost, https://www.marktechpost.com/2026/10/14/fish-audio-drama-3-open-weight-tts-model/}
}Credit the source
Original reporting: MarkTechPost, “Fish Audio Releases Drama 3, a 35B Open-Weight MoE Text-to-Speech Model”, https://www.marktechpost.com/2026/10/14/fish-audio-drama-3-open-weight-tts-model/(opens in a new tab)
Follow this story
Get Artificial Intelligence in your inbox — new stories in 40 words or fewer, once a week.
Subscribe