Fish Audio officially open-sources Drama 3 model weights
BIBIBI
AT A GLANCE
Fish Audio announces the open-sourcing of Drama 3 weights and code, stating that it supports 83 languages and fine-grained voice control.
Article
Fish Audio open-sources Drama 3: based on Qwen3.5-35B-A3B, focused on fine-grained voice control
Beating AI news,AI voice company Fish Audio has officially open-sourced the Drama 3 model weights, allowing developers to download and deploy it independently from Hugging Face. Drama 3 released a preview version as early as September and has now fully opened access to the web, App, and API, while also making the model weights, inference code, and technical report public.
Drama 3 is built on Qwen3.5-35B-A3B language model, with a total of 35000000000 parameters; each generation activates only about 3000000000 parameters, and it supports 83 languages. It focuses on fine-grained control of vocal performance: users can directly describe emotions in natural language, making a line shift from calm to choked with emotion, insert laughter at specified positions, and even have up to 16 characters take turns speaking within the same text.
In Artificial Analysis’s English pronunciation accuracy test, Drama 3 ranked first with a score of 93.2%, higher than the previous-generation S2.1 Pro’s 88.1%。
Drama 3 uses the Fish Audio Community License. Research and non-commercial use are free; individuals, companies, and their affiliates with combined annual revenue below 500000 USD may use it commercially for free within the scope of the license. However, providing the model as a voice generation, cloning, or inference API service for others still requires separate authorization.
Original link https://m.theblockbeats.info/flash/371736
Key points
01
The original text states that Fish Audio has officially open-sourced the Drama 3 model weights, allowing developers to download and deploy them independently from Hugging Face.
02
The original text states that access has been opened to the web, App, and API, while the model weights, inference code, and technical report have also been made public.
03
Drama 3 built on Qwen3.5-35B-A3B with a total of 35000000000 parameters, with approximately 3000000000 parameters activated per generation.
04
The original text states that Drama 3 supports 83 languages, and supports describing emotions in natural language, inserting laughter at specified positions, and having up to 16 characters take turns speaking.
05
The original text states that, in Artificial Analysis’s English pronunciation accuracy test, Drama 3 ranked first with 93.2%, while S2.1 Pro scored 88.1%。
06
Drama 3 uses the Fish Audio Community License; the original text states that research and non-commercial use are free, and that when combined annual revenue of an individual, company, and its affiliates is below 500000 USD, it may be used commercially for free within the scope of the license.
07
The original text states that providing the model as a voice generation, cloning, or inference API service for others still requires separate authorization.
AI-assisted interpretation
The following is analysis, separate from reported facts. Verify important claims independently.
Drama 3 is a model that can generate and control vocal performance. According to the original text, developers can use not only the web, App, or API, but also download the weights and inference code for independent deployment. Its focus is allowing users to specify speaking emotions, laughter positions, and turn-taking among multiple characters in text.
Why it matters to readers
Making the weights and inference code public may give developers more room for deployment and customization; 83 languages, emotion control, and having up to 16 characters take turns speaking may expand its uses in multilingual and multi-character voice content. Commercial use remains subject to license and additional authorization conditions.
Beginners can think of it as a voice model that can “act out” lines according to textual instructions. Before trying it, confirm the download and deployment requirements; if it is used in a commercial product or to provide voice API services to others, carefully review the license conditions as well.
Risks and unknowns
The provided materials do not include links to an official Fish Audio announcement, Hugging Face model page, code repository, or technical report, so the open-source content and actual availability for download cannot be cross-verified based on these materials.
93.2%'s test scores, rankings, and comparison with S2.1 Pro lack a test page and specific evaluation conditions.
The original license text was not provided; the scope of free commercial use, the method of calculating revenue, and the specific requirements for “obtaining authorization separately” still need to be verified.
Related Developments
Loading event timeline…
The provided materials do not specify the hardware, costs, performance, or usage limitations required for actual deployment.
Related concepts
model weights
This term is not in the glossary yet. Browse related concepts in the glossary.
