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

What it does

Fish Audio launched Drama 3, a text-to-speech (TTS) model built on a mixture-of-experts (MoE) architecture. It has 35 billion parameters total but activates only about 3 billion at once. Drama 3 supports plain-language direction inputs and can generate speech in 83 languages. Fish Audio is releasing the model with open weights under a revenue-capped community license, enabling wider access to a large-scale TTS model.

Why it matters

Drama 3 lowers the barrier to high-quality multilingual TTS with a versatile model that does not rely on proprietary weights or tightly controlled APIs. The MoE design allows it to scale parameter count while keeping compute costs manageable, which helps balance quality with efficiency. Covering 83 languages expands opportunities for global voice applications in assistants, audiobooks, and accessibility tools. The open-weight release under a capped-license challenges the increasingly closed nature of high-end TTS offerings, potentially shifting power toward smaller developers and open source projects.

Who it is for

Developers and startups focusing on voice-enabled products gain a powerful, flexible TTS model they can fine-tune or build on without negotiating costly licenses. Enterprises requiring multilingual support for customer service, content localization, or accessibility improvements can evaluate Drama 3 as a cost-effective alternative to commercial APIs. Researchers and hobbyists benefit from direct access to weights for experimentation and innovation in speech synthesis.

The catch

Despite the large parameter scale, only about 3 billion are active at once, which may limit peak quality compared to fully dense models of similar size. The community license includes revenue caps, so commercial users must monitor usage and growth carefully to avoid licensing violations. The complexity of MoE models requires specialized infrastructure and expertise, which could raise integration costs for smaller teams.

What to watch next

How Fish Audio’s open licensing affects the competitive landscape among TTS providers will be key. Watch for adoption rates in startups and research groups that typically avoid closed-source models. The practical impact on cost and quality trade-offs in multilingual TTS services will influence whether open-weight MoE models become standard or stay niche. Finally, monitor if other vendors follow with similar open models or adjust pricing in response.

AI Quick Briefs Editorial Desk