Fish Audio launched Drama 3 on October 13 across its web service, mobile apps and API. The model supports expressive text-to-speech, voice cloning and natural-language direction for multi-character scenes.
Its model card describes a mixture-of-experts backbone with 35B parameters and about 3B active per token, with tags for up to 16 speakers. Downloadable weights and serving documentation support self-hosted evaluation.
The actual license file is Fish Audio Community License v1.0. Commercial eligibility includes a $500,000 annual-revenue threshold, while hosted TTS or cloning services require separate permission. Voice cloning also requires consent; “open weights” does not mean unrestricted use.
Sources: Fish Audio launch post · Official model card · Actual license file