How to Run Zonos 3 Mini Locally for 43-Language TTS
Run Zyphra's Zonos 3 Mini text-to-speech locally: install steps, GPU memory needs, Apache 2 license, 43 languages and how voice cloning actually sounds.
For Marketing, Engineering and IT

How do you run Zonos 3 Mini locally?
Clone Zyphra's ZONOS3-Mini repository from GitHub, install its dependencies with the UV package manager, and start the included server. In one hands-on test, the model used close to 25 GB of GPU memory on a 48 GB card, so a GPU with about 24 GB is a sensible target. You can also run it in a WebGPU browser, behind SGLang-Omni, or on Apple Silicon and CPU through Zyphra's other runtimes.
TL;DR
- Zonos 3 Mini is an open-source text-to-speech model from Zyphra, released under the Apache 2.0 license, which allows commercial use.
- It supports 43 languages and can clone a voice from a short reference clip.
- The model has 800 million parameters in total but only about 100 million are active for each token, which is why it is fast.
- In one hands-on test it used close to 25 GB of VRAM, so plan for a 24 GB GPU, though other runtimes may need much less.
- In testing, European and Indian languages sounded good, while cloning was decent but a little plasticky in places, notably in Chinese.
- You can try it with no install at all through a WebGPU browser studio before committing to a server.
What is Zonos 3 Mini?
Zonos 3 Mini is a text-to-speech (TTS) model. You give it written text, and it returns spoken audio. Zyphra describes it as compact and open source, with natural speech and voice cloning from only 100 million active parameters.
Two features matter most for everyday work:
- Voice cloning. Give it a reference audio clip and it speaks new text in that voice. Zyphra's documentation says the clip can be up to 60 seconds.
- Voice design and control. You can shape a character by age, accent and vocal texture, and steer emotion, pace and delivery. The model also takes controls for loudness, audio bandwidth and speaking rate.
The weights are released under Apache 2.0. In plain terms, you can use it in a commercial product without paying a license fee. Still, read the license text yourself before shipping anything.
How much GPU memory does Zonos 3 Mini need?
The only measurement in the hands-on test came from a machine with a 48 GB GPU. There, the model was using close to 25 GB while serving. The reviewer's conclusion was that it should run on a GPU of around 24 GB.
Treat that as a practical target, not a hard floor. The weights are only 800 million parameters stored in bf16 (a 16-bit number format), which is small. So the 25 GB figure probably reflects how the server allocates memory, but that is my inference, not something the sources state. Zyphra also ships runtimes meant for lighter hardware:
| Runtime | Hardware |
|---|---|
| CUDA server | NVIDIA GPUs |
| MLX server | Apple Silicon |
| C++/ggml server and CLI | CPU, CUDA, Metal or Vulkan |
| WebGPU build | Your browser, using your own GPU |
If you do not have a 24 GB card, start with the browser or ggml options and see whether the speed is acceptable for your use.
How do you install and serve it, step by step?
The workflow shown in the hands-on test has three parts. The exact commands live in the repository's README, so copy them from there rather than from a blog post.
- Clone the repository. The code is at github.com/Zyphra/ZONOS3-Mini.
- Install prerequisites with UV. UV is a fast Python package manager. Follow the README's install commands.
- Serve it. Start the server. It runs the Zonos Voice Studio, a web interface you open in your browser. In the test it ran on the reviewer's own server.
Zyphra's native runtimes share one HTTP API, so what you build against one server carries over to another.
The SGLang alternative
If you want an API your existing tools can call, Zyphra's model card documents serving through SGLang-Omni on NVIDIA GPUs. After installing SGLang-Omni, the start command is:
sgl-omni serve --model-path Zyphra/ZONOS3-Mini --port 8000
Check that curl http://localhost:8000/health reports healthy. Then generate speech through an OpenAI-compatible endpoint:
curl -X POST http://localhost:8000/v1/audio/speech \
-H "Content-Type: application/json" \
-d '{"input": "Hello from ZONOS3-Mini!"}' \
--output output.wav
To clone a voice, add a references field pointing to your audio clip. No transcript of that clip is needed. For streaming, set "stream": true and "response_format": "pcm", and you get 44.1 kHz audio as it is generated.
Can you try it without installing anything?
Yes. The model card lists two browser options that run entirely on your own GPU through WebGPU, with no server and no install:
- Zonos Voice Studio, the full studio.
- Lite, a quick voice cloning page where you pick, upload or record a voice and speak with it.
The first visit downloads the model, and the browser caches it. You need a browser that supports WebGPU. This is the fastest way to judge quality for your own language and text before setting up a server.
Which languages does it support?
Zonos 3 Mini supports 43 languages, and Zyphra's list covers a wide spread:
- Europe: English, French, Spanish, German, Italian, Portuguese, Dutch, Greek, Swedish, Danish, Norwegian, Finnish, Polish, Czech, Slovak, Hungarian, Romanian, Bulgarian, Russian, Ukrainian.
- Middle East and Caucasus: Turkish, Arabic, Hebrew, Georgian.
- East and Southeast Asia: Mandarin Chinese, Japanese, Korean, Vietnamese, Indonesian, Malay, Thai, Tagalog.
- South Asia: Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Odia, Urdu.
The repository includes 27 default voices, grouped by language. One detail worth knowing: the model's text normalization (turning "$1,249.99" or "8:45 a.m." into spoken words) is built in for nine languages, namely English, Mandarin, Japanese, Korean, German, Spanish, French, Italian and Portuguese. Other languages pass through as written, so numbers and abbreviations may need cleanup in your own text first.
How good are the speech and voice cloning?
Here is what the hands-on test showed. These are one reviewer's impressions on a handful of samples, not a benchmark.
Stock voices. A deliberately hard English paragraph with dates, a price, a weight, an email address and tongue twisters came out as "not a bad rendering," and generation ran in real time. French and Italian also sounded good. Hindi, Korean, Telugu, Spanish and Portuguese were described as "not bad," and Portuguese as "not perfect."
Voice cloning. The reviewer uploaded a recording of his own voice and had it speak Telugu, which worked and was judged "not bad." An English passage cloned from the same recording sounded weaker to him. In Chinese, the cloned voice kept the texture of the original, but the pronunciation and delivery were still a bit "plasticky."
So the pattern is: good default voices across many languages, cloning that captures a voice's character, and some artificial delivery that varies by language and sample. Test your own language with your own clip before you depend on it.
Zyphra's own numbers point to speed. On a single H100 80 GB, it reports 80 milliseconds to first audio with 8 concurrent requests, and a peak of 457 seconds of audio generated per second. Those are Zyphra's figures on datacenter hardware, so expect different results on a 24 GB card.
Why is it small and fast?
The model is a mixture of experts. It holds 64 specialist sub-networks, and for each step it picks two. That is how 800 million total parameters turns into about 100 million doing work per token.
The model predicts audio as a stream of compact codes, 15 frames per second, with 10 codes per frame. One code captures what is being said, drawn from a speech model, and nine hold the fine acoustic detail. A decoder turns each frame into 44.1 kHz audio as soon as it is produced, which is why streaming works. A single generation can run up to 10 minutes of audio.
What does this mean for you?
If you make training videos, podcasts or localized content, a local model means you can narrate scripts in many languages without per-minute fees or sending scripts to a third party. Draft in your own language, generate a rough voiceover, then have a native speaker check it before publishing. The test suggests quality is good enough for drafts, and sometimes for final use, but not uniformly.
If you work in IT or run a small team's tools, the Apache 2.0 license and OpenAI-compatible endpoint make it easy to slot into existing workflows. Start with the browser studio, then a 24 GB machine for a shared server.
If you plan to clone voices, get written permission from the person whose voice it is. Cloning your own voice for your own content is straightforward. Cloning a colleague or a public figure is a legal and trust problem, whatever the license says.
Could you build a small tool around a voice model like this?
Probably, and the useful part is rarely the model itself. It is the workflow around it: a form where a colleague pastes a script, picks a language and voice, and gets a file back, with a history of what was generated. Remy builds that kind of internal app from a plain-English description, with login, a database and hosting included, and its catalog includes AI voice and audio models. Describe the request form, the approval step and the storage you want, and you get a deployed app you own instead of another subscription. Remy can call a self-hosted model like Zonos, with built-in support for registering custom or self-hosted models or reaching local models over a secure tunnel.
Frequently Asked Questions
Is Zonos 3 Mini free for commercial use?
The weights are released under the Apache 2.0 license, which permits commercial use. You still need to follow the license terms, and you are responsible for having rights to any voice you clone.
Do I need an NVIDIA GPU?
Not necessarily. The CUDA and SGLang-Omni routes are for NVIDIA GPUs, but Zyphra also lists an Apple Silicon (MLX) server, a C++/ggml server that runs on CPU, Metal or Vulkan, and a WebGPU browser build.
How long a reference clip does voice cloning need?
Zyphra's documentation says the reference clip can be up to 60 seconds. The hands-on test used a short uploaded recording of the reviewer's own voice. Cleaner audio with a single speaker is the safer choice.
Is there a hosted version if I do not want to run it myself?
Yes. Zyphra lists a hosted version on Zyphra Cloud. Running it locally keeps your text and audio on your own machine, while the hosted version saves you the setup.
How we source, write, and correct our articles. Editorial standards



