ZONOS3-Mini: open TTS in 43 languages that runs in a browser
Zyphra released ZONOS3-Mini on 9 November: an 800M mixture-of-experts text-to-speech model with 100M active parameters, 43 languages, voice cloning and Apache 2.0 weights.
Zyphra, a San Francisco AI lab, released ZONOS3-Mini on 9 November 2026 under the Apache 2.0 license. It is a text-to-speech model that speaks 43 languages and can clone a voice from a reference clip of up to 60 seconds.
What is new
ZONOS3-Mini is a mixture-of-experts model with 800 million parameters and about 100 million active per token. Zyphra designed it to run on phones and laptops, including inside a browser tab through WebGPU. The language model predicts audio frames at 15 per second in Zyphra’s own codec, ZAC-1, and a causal decoder renders 44.1 kHz audio as it streams. One generation can last up to 10 minutes. Besides voice cloning, it has controls for age, gender, accent and voice texture, steering for pace, emotion and delivery down to single words, and an edit model that rewrites part of a finished sentence.
Zyphra reports a median time to first audio of 80 ms with 8 concurrent requests on one H100 GPU. In a browser it reports under 570 ms on a Pixel 9 and under 150 ms on an M3 MacBook Air. On its own Zyphra TTS Capability Index, built from its ZTTS2-Eval benchmark of 55 language subsets, ZONOS3-Mini ranks first among open models, ahead of Fish Audio S2 Pro and MOSS TTS v1.5. The benchmark and the index are both Zyphra’s.
Running it
The language-model weights are 1.8 GB in BF16, and the PyTorch weights that the CUDA server needs, including the audio codec, are about 9 GB. The Hugging Face repository is about 31 GB because it also holds training files, and each runtime needs only specific files. Zyphra ships runtimes for NVIDIA GPUs with SGLang-Omni, for Apple Silicon with MLX, for CPU, CUDA, Metal and Vulkan with a C++ server, and for browsers with WebGPU. It also sells a hosted version on Zyphra Cloud.
License
The weights use Apache 2.0, so there is no revenue cap and no ban on offering the model as a hosted service, unlike Drama 3 from Fish Audio, which has both. Apache 2.0 does not change the rules on consent, though: cloning a real person’s voice without permission can be illegal or harmful. Text normalization covers only nine languages, so numbers and abbreviations in the others may be read badly.
Why it matters
ZONOS3-Mini adds a cloning model with plain terms that can run in a browser, in a field where Kokoro offers fixed voices in 8 languages and Pocket TTS clones in 6. It is also brand new, so it has little track record, and its ranking is Zyphra’s own. Listen to your language and your voices before you choose.
For more help, read how to choose an open-source text-to-speech model, or see ZONOS3-Mini vs Drama 3 for a side-by-side comparison.
Sources
- ZONOS3-Mini (Zyphra blog, 9 Nov 2026) zyphra.com
- ZONOS3-Mini model card (Hugging Face) huggingface.co
- ZONOS3-Mini runtimes and README (GitHub) github.com
- ZTTS2-Eval benchmark (Zyphra) zyphra.github.io
Models in this story
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| ZONOS3-Mini Zyphra | 800M 100M active | – | Apache 2.0 (Permissive) | 10 GB Laptop or phone | |
| Drama 3 Fish Audio | 35.9B 3B active | – | Fish Audio Community 1.0 (Conditional) | 25 GB Gaming GPU | |
| Kokoro hexgrad | 82M | – | Apache 2.0 (Permissive) | 0.38 GB Laptop or phone | |
| Pocket TTS Kyutai | 18M – 100M | – | CC BY 4.0 (Permissive) | 0.127 GB Laptop or phone |