ZONOS3-Mini vs Drama 3: which new open TTS model?
ZONOS3-Mini is an 800M Apache 2.0 model for 43 languages that runs in a browser. Drama 3 is a 36B model for 83 languages, free only below US$500K revenue.
Verdict. Choose by size and license. ZONOS3-Mini has 800M parameters with about 100M active, 43 languages, controls for accent, age, emotion and emphasis on single words, and plain Apache 2.0 weights, and Zyphra ships runtimes that run it in a browser, on Apple Silicon or on a CPU. Drama 3 has 36B parameters with about 3B active, 83 languages, directions written into the text such as [whispering], and up to 16 speaker tags in one script, but its main checkpoint is 41.5 GB and the Fish Audio Community License is free for commercial use only below US$500,000 of yearly revenue and never covers a hosted voice service. Each vendor's quality claim comes from its own benchmark: Zyphra ranks ZONOS3-Mini first among open models on its own index, and Fish Audio reports 93.2% on Artificial Analysis' Pronunciation Robustness benchmark, so the two claims cannot be compared. Both clone voices, so get the speaker's consent first.
Specifications
| Specifications | ZONOS3-Mini | Drama 3 |
|---|---|---|
| Developer | Zyphra | Fish Audio |
| Released | November 9, 2026 | October 13, 2026 |
| Total parameters | 800M (ZONOS3-Mini) | 35.9B (Drama 3) |
| Active parameters | 100M | 3B |
| Architecture | Mixture of experts | Mixture of experts |
| Context window | – | – |
| Input | Text, Audio | Text, Audio |
| License | Apache 2.0 (Permissive) | Fish Audio Community 1.0 (Conditional) |
| Memory for the smallest variant (4-bit) | 10 GB (ZONOS3-Mini) | 25 GB (Drama 3 INT4 (hybrid W4A16, unofficial quantization)) |
Performance claims come from the developer and have not been independently verified.
Pick ZONOS3-Mini if
- You want plain Apache 2.0 weights with no revenue cap, including for a hosted voice service.
- You need to run on a laptop, a phone or in a browser: Zyphra ships WebGPU, MLX and C++ runtimes, and the language-model weights are 1.8 GB in BF16.
- You want fine control over delivery: accent, age, emotion and emphasis on single words, plus an edit model that rewrites part of a finished sentence.
- You can live with 43 languages, and with text normalization for only nine of them.
Pick Drama 3 if
- You need 83 languages, or acted scenes with up to 16 speakers in one script.
- You want directions in the text, such as [whispering] or [laughs], which can be written in a different language from the line itself.
- You have a GPU with about 48 GB for the FP8 build, or about 25 GB for the unofficial INT4 build, by our estimate.
- Your business earns under US$500,000 a year and you use the model for your own content or product, not as a voice service for others.