Zyphra · 2026-11-09 · notable
ZONOS3-Mini — Zyphra's open TTS model clones voices in real time
In short
ZONOS3-Mini is Zyphra's open-weight text-to-speech model: an 800M-parameter MoE with about 100M active, Apache 2.0, 43 languages and zero-shot voice cloning. It runs in real time on laptops, phones and in the browser.
Zyphra · 2026-11-09 · 3 sources checked
An Apache-2.0 speech model small enough to clone a voice in real time on a phone or in a browser tab.
Key specs
| Active params | ~100M |
|---|---|
| Languages | 43 |
What is it?
ZONOS3-Mini is a compact open-weight text-to-speech model from Zyphra that clones a voice from a short audio sample and speaks 43 languages. Beyond cloning, it can design a voice from age, accent and vocal texture, and it lets you steer emotion, pace and delivery down to single words. The weights are on Hugging Face under Apache 2.0.
How does it work?
Under the hood the model is a sparse mixture of experts: 800M parameters in total, but only about 100M are active for each token. It predicts 15 Hz audio tokens from Zyphra's own ZAC-1 codec and outputs 44.1 kHz audio, up to 10 minutes per generation. Zyphra trained it on 112 NVIDIA H100s in 2.5 days and post-trained it with DPO and then GRPO.
Why does it matter?
Local, real-time voice cloning used to mean a GPU server. Zyphra reports time to first audio under 150 ms on an M3 MacBook Air and under 570 ms in a browser on a Pixel 9, and on an H100 it serves 80 ms first audio with 8 concurrent requests. That makes on-device voice agents and offline narration practical, with day-0 support in SGLang, vLLM, Ollama and llama.cpp.
Who is it for?
developers building voice agents and speech apps
Try it
ollama run hf.co/Zyphra/ZONOS3-Mini