Zyphra Releases ZONOS3-Mini, an Open-Source TTS Model That Clones Voices and Runs in the Browser
Zyphra has released ZONOS3-Mini under the Apache 2.0 license, an 800-million-parameter text-to-speech model with 100 million active parameters capable of cloning voices from an audio sample and adjusting accent, age, emotion, or speaking rate down to the individual word. It runs in real time on laptops, phones, and directly inside the browser.
Cloning a voice in seconds no longer requires a server. Zyphra, a California-based lab known for its open models, has released ZONOS3-Mini, a text-to-speech engine licensed under Apache 2.0 that fits on a phone and runs directly inside the browser.
Small in Compute, Big on Parameters
ZONOS3-Mini is a mixture-of-experts model: 800 million total parameters, but only around 100 million active at any given moment. Zyphra explains that it lowered resource demands across multiple fronts: a new proprietary audio codec (ZAC-1) that slices audio into 15 frames per second instead of the 86 used in ZONOS2, a hybrid architecture borrowed from large language models, and two post-training stages to improve robustness. Training took 2.5 days across 112 H100 GPUs, according to the company.
The advertised result: time-to-first-audio drops below 570 ms inside the browser on a Pixel 9 and below 150 ms on an M3 MacBook Air. On its Hugging Face page, Zyphra provides a voice studio that runs entirely on the user’s GPU via WebGPU, along with inference backends for Nvidia, Apple Silicon, and CPU.
Clone, Design, Direct
The model clones a voice from an audio clip of up to one minute, without needing a transcription. It also allows users to “compose” a voice by tweaking age, gender, accent, and texture (smooth, raspy), and then direct the performance: emotion, speaking rate, projection, expressiveness, or even whispering just a single part of a sentence. A dedicated editing model can modify already generated phrases. Text normalization covers nine languages, including French, and 27 default voices are supplied out of the box.
In parallel, Zyphra released a benchmark, ZTTS2-Eval, and an index that ranks ZONOS3-Mini at the top of open models, ahead of Fish S2 Pro and MOSS TTS v1.5. This evaluation was designed by the model developer itself.
Our take
For creators, game studios, and voice-assistant developers, an open model with this reported quality, a permissive license, and the ability to run offline alters the calculus against pay-per-minute services. That same accessibility inevitably raises concerns around non-consensual cloning: the Apache 2.0 license imposes no usage restrictions, and nothing in the announcement prevents cloning someone else’s voice from a short snippet. On the quality front, the comparative figures stem entirely from Zyphra’s internal benchmark; independent listening tests, particularly in French, will show whether the model truly delivers on its promise.
Sources
- Zyphra — ZONOS3-Mini (November 9, 2026) Primary source
- Hugging Face — Zyphra/ZONOS3-Mini Primary source
A primary source comes directly from the company, lab or organization concerned. A secondary source reports the information without being its author: we flag it as such. Spotted an error? See our corrections policy.
