Fish Audio releases Drama 3, a 35B open-weight text-to-speech model

Voice AI company Fish Audio has released Drama 3, a new text-to-speech model announced on October 14, 2026. The model has 35 billion parameters in total and uses a mixture-of-experts (MoE) design. Fish Audio is making its weights publicly available.

According to the release details, Drama 3 activates roughly 3 billion parameters at a time. In an MoE system, only a subset of specialised sub-networks handles each input, which lets a model keep a large overall capacity while keeping the compute needed to run it comparatively modest.

Directed in plain language

A central feature of the model is that users can steer the voice output with ordinary written instructions rather than technical settings, describing the tone, emotion or delivery they want. Drama 3 is said to support 83 languages.

The name and the emphasis on natural-language direction point to a focus on expressive, performance-style speech. The publicly available summary, however, does not include benchmark comparisons on audio quality or a breakdown of how well each language is supported.

Licensing terms

The weights are distributed under a community license with a revenue cap. Licenses of this kind typically let individuals, researchers and smaller companies use a model freely, while requiring businesses above a set revenue threshold to negotiate separate commercial terms. The exact threshold Fish Audio has chosen was not specified in the available information.

Open weights allow developers to run the model on their own hardware, fine-tune it and build it into products without relying on a hosted service. Since much of the text-to-speech market is served through closed, API-based offerings, an open model at this scale gives teams working on voiceover, games and accessibility tools another option to evaluate.