Zyphra releases ZONOS3 Mini, a free voice model with cloning in 43 languages

Zyphra released ZONOS3 Mini, an open text to speech model that clones voices in 43 languages and runs on a laptop or in a browser tab. Here is what it costs and where it fits.

ZONOS3 Mini title card from Zyphra, white lettering over a sound wave

You need a 30 second voiceover for a product video. You could record it yourself in the car, pay a voice actor, or rent a monthly plan from a text to speech service. This week Zyphra added a fourth option: a free, open text to speech model called ZONOS3 Mini that runs on your own laptop, and even inside a browser tab.

Zyphra published the model on Hugging Face under the Apache 2.0 license. It speaks 43 languages, clones a voice from a short audio clip, and lets you shape age, accent and vocal texture to design a new character voice. Zyphra says it runs in real time on phones and laptops.

Key facts

  • Open weights under Apache 2.0, free to download and run.
  • 43 languages, voice cloning from a reference clip, up to 10 minutes of audio per generation.
  • About 100 million active parameters out of 800 million in total, small enough for WebGPU in the browser.

What ZONOS3 Mini does

You type text and pick a voice, and the model returns 44.1 kHz audio. Zyphra lists three ways to choose that voice:

  • Clone a voice. Feed it a reference clip. Zyphra's example uses 5 to 10 seconds, and the model accepts up to 60 seconds. You do not need a transcript of the clip.
  • Design a voice. Set age, accent and texture to build a character from scratch.
  • Control delivery. Zyphra says you can steer emotion, pace and delivery down to single words. The technical table lists loudness, bandwidth and speaking rate as the built in controls.

On server hardware Zyphra reports a median of 80 ms before the first audio arrives, measured with 8 requests at once on one NVIDIA H100.

Who can use it and what it costs

Anyone can download the weights from Hugging Face. Zyphra publishes runtimes for NVIDIA GPUs, Apple Silicon (MLX), a C++ build for CPU, and a WebGPU build that runs in a browser with nothing to install. The model card also lists local apps such as LM Studio, Jan and Ollama.

Zyphra also hosts the model on Zyphra Cloud for people who would rather call an API. The model card does not show a price for that hosted version.

OptionWhat you needCost shown by Zyphra
Browser playgroundA browser with WebGPUFree, runs on your own GPU
Local app or scriptLaptop or desktop, some setupFree under Apache 2.0
Zyphra CloudAn accountNo price on the model card

What it means for a small business

Voiceovers show up in more places than they used to: Reels, product demos, explainer videos, on hold messages, training clips for new staff. A free model that runs on your own computer means you can make draft voiceovers without a subscription and without uploading scripts to a third party service.

Two cautions. Clone only your own voice or a voice you have written permission to use; the model card has no section on consent or watermarking, so that judgment sits with you. And open models take more setup than a paid app. If you have never run a local model, start with the browser playground before installing anything.

Our take

Based on the announcement, not on hands on testing: ZONOS3 Mini is worth a try if you pay for a voice tool only for occasional voiceovers. The Apache 2.0 license allows commercial use, and the browser version removes most of the setup. Keep your paid tool for now if you depend on a polished editor, a big voice library or support. Listen to the samples in your own language first, since quality across 43 languages will not be even.

Frequently asked questions

Is ZONOS3 Mini free for commercial use?

Zyphra released the weights under Apache 2.0, a license that permits commercial use. Check the license text on the model page before you rely on it.

Can ZONOS3 Mini run without a powerful computer?

Zyphra says it runs in real time on phones and laptops, and it offers a WebGPU version that runs in a browser. You need a browser that supports WebGPU.

How long a sample does voice cloning need?

Zyphra's example uses a 5 to 10 second reference clip, and the model accepts clips up to 60 seconds, with no transcript needed.

Sources

More AI news

All AI news