Zyphra's ZONOS3-Mini is an Apache-2.0 TTS model that clones voices in a browser

Zyphra has released ZONOS3-Mini, an open text-to-speech model with voice cloning. The weights are on Hugging Face under Apache-2.0, and Zyphra’s runtimes are on GitHub under MIT.

What it is

ZONOS3-Mini is a sparse mixture-of-experts model with 800 million parameters in total and about 100 million active per token. Zyphra’s release post says it was trained on 112 H100s in two and a half days. Its predecessor, ZONOS2, used 8 billion parameters and an 86 Hz audio codec, which Zyphra says limited where it could run; ZONOS3-Mini uses Zyphra’s own codec at 15 Hz.

From the model card:

• Output: 44.1 kHz audio, up to 10 minutes per generation, streamed as it is produced.
• Cloning: from a reference clip of up to 60 seconds, or from a stored speaker embedding. 27 default voices ship with it.
• Languages: text normalisation for English, Mandarin, Japanese, Korean, German, Spanish, French, Italian and Portuguese; other languages pass through as written.
• Control: loudness, bandwidth and speaking-rate settings, plus emotion, pace and delivery steering that can target single words. An edit model can change a line after generation.

Where it runs

This is the practical part. ZONOS3-Mini runs in the browser through WebGPU, with no install, and Zyphra offers a local playground. Zyphra says time to first audio is under 570 ms on a Pixel 9 and under 150 ms on an M3 MacBook Air. There are also CUDA, Apple MLX and C++/ggml servers that load GGUF files, and day-zero support in SGLang-Omni behind an OpenAI-compatible /v1/audio/speech endpoint:

sgl-omni serve --model-path Zyphra/ZONOS3-Mini --port 8000

On one H100, Zyphra reports 80 ms median time to first audio with 8 concurrent requests and a peak of 457 seconds of audio per second. A hosted version is on Zyphra Cloud.

How good is it

Zyphra also released a new benchmark, ZTTS2-Eval, covering 35 languages of read speech and 20 of in-the-wild recordings, and a capability index built on it. On that index, Zyphra ranks ZONOS3-Mini first among open models, ahead of Fish S2 Pro and MOSS TTS v1.5. That is Zyphra’s test of its own model; it is not yet on independent boards such as Artificial Analysis.

We found no consent or misuse checks for cloning described in the model card or release post. With Apache-2.0 weights, that responsibility sits with whoever deploys it.

Why this matters

Our open-weight voice review found a trade-off: Kokoro is tiny but cannot clone, Chatterbox clones under MIT, and Fish Audio Drama 3 sounds best but needs data-centre GPUs and a revenue-capped licence. ZONOS3-Mini claims cloning, Apache-2.0 and laptop or browser speed in one model, so it is worth a side-by-side test on your own scripts. Our open speech models guide has the wider field. If you clone voices with it, keep consent records yourself, and for live agents pair it with the rest of the voice pipeline.