Fish Audio Releases Drama 3, a 35B Open-Weight MoE Text-to-Speech Model

Fish Audio has released Drama 3, a large-scale text-to-speech model built on a mixture-of-experts (MoE) architecture. The model totals 35 billion parameters but activates only about 3 billion per inference, which should keep compute costs manageable while retaining the capacity of a much larger system.

The model accepts plain-language direction rather than rigid markup, allowing users to describe tone, emotion, and pacing in natural language. It supports 83 languages, making it one of the more broadly multilingual open TTS options available.

Weights are distributed under a community license with a revenue cap, meaning organizations exceeding a certain revenue threshold must negotiate commercial terms. This approach lets hobbyists and smaller teams access the model freely while reserving large-scale commercial use for paid arrangements.

Open-weight TTS models have been a growing trend, but few combine MoE efficiency with wide language coverage. Drama 3's architecture could influence how other teams balance model size, inference cost, and multilingual support in speech generation.

This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.