Fish Audio has released Drama 3, a text-to-speech model built on a 35B mixture-of-experts architecture. Although the total parameter count is large, only about 3B parameters are active during use, which can keep inference more efficient than a dense model of similar size.
The model is designed to follow plain-language direction, so users can describe delivery or style without needing low-level control codes. It covers 83 languages, broadening its utility for multilingual voice generation.
Drama 3 ships with open weights under a community license that includes a revenue cap. That means developers can access and modify the model for many projects, but commercial use above a certain income threshold would require a different arrangement. The release adds another option in the growing field of open-weight speech models, though the source does not compare it directly with rivals.