How to Run Zonos 3 Mini TTS Locally on Your Own GPU

What is Zonos 3 Mini?

Zonos 3 Mini is an open-weight text-to-speech model from Zyphra, released under the Apache 2.0 license, which means it can be used commercially without restriction. It converts written text into spoken audio and can clone a voice from a short reference clip. The model supports 43 languages and is built as a mixture-of-experts (MoE) system with 800 million total parameters, but only around 100 million are active per token because the model routes each step through a couple of experts chosen from a pool of 64. That small active-parameter count is why it can run comfortably on consumer and prosumer GPUs despite having a fairly large total parameter budget.

TL;DR

Zonos 3 Mini is Zyphra’s Apache 2.0 licensed TTS model that turns text into natural speech and supports voice cloning from a short audio sample.
The model uses a mixture-of-experts architecture with 800 million total parameters but only about 100 million active per token, picking from 64 experts.
It was demonstrated running on a local server with 48 GB of VRAM available, consuming roughly 25 GB, suggesting it’s workable on a 24 GB GPU.
Setup involves git cloning the repository, using the UV package manager to install dependencies, and serving the model locally through Zyphra’s Voice Studio interface.
The model covers 43 languages, including English, Hindi, Telugu, French, Italian, Korean, Spanish, Portuguese, Chinese, and Irish-accented English, with voice cloning quality holding up reasonably well across most of them.
Under the hood, a codec system compresses audio to 15 frames per second, using one code book for semantic content (via a frozen W2V-BERT style speech model) and nine more code books for fine acoustic detail.
Output quality was described as solid for a TTS model but not flawless, with some languages sounding more natural than others and occasional “plasticky” delivery in certain tests.

Developers have their tools.

Enterprises have Remy.

How do you install and run Zonos 3 Mini locally?

The setup process shown for Zonos 3 Mini follows a straightforward pattern common to many open-weight model releases:

Clone the repository. Use git clone to pull down Zyphra’s Zonos 3 Mini codebase to your local machine.
Install dependencies with UV. Rather than relying on pip directly, the setup uses the UV package manager to handle prerequisites, which tends to be faster and more reliable for resolving Python dependency trees.
Serve the model. Once dependencies are installed, you launch the serving command, which spins up Zyphra’s Voice Studio interface locally. This gives you a web UI where you can type or paste text, pick a language and voice, and generate audio.

No cloud account or API key is needed for inference since everything runs on your own hardware. The main friction point is making sure your GPU has enough VRAM headroom, and that your environment has the right CUDA and Python versions UV expects.

What are the VRAM and hardware requirements?

In testing, the model was served on a machine with 48 GB of VRAM, and the running process consumed close to 25 GB. That suggests a 24 GB GPU (think RTX 4090, RTX 3090, or similar prosumer cards) should be enough to run Zonos 3 Mini, though the exact margin will depend on batch size, sequence length, and whatever else is sharing the GPU at the time.

This VRAM footprint is noticeably higher than the model’s active-parameter count alone would imply. An MoE model with 100 million active parameters per token sounds small, but all 64 experts (and the full 800 million parameters) typically need to sit in memory even though only a fraction fire per step. Add in the audio codec, the text embedding pipeline, and any KV-cache-style buffering for streaming generation, and the real-world VRAM number climbs well past what the “active parameters” figure alone suggests.

If you’re working with less VRAM, options include running on a rented GPU instance, waiting for quantized community variants to appear, or testing on shorter text inputs to reduce memory pressure during generation.

How does the Zonos 3 Mini architecture work?

The model’s pipeline has two main input paths that feed into a shared generation core:

Speaker pipeline. A reference audio clip is processed into either a raw audio codec representation or a single speaker embedding vector. This tells the model whose voice to imitate.

Text pipeline. Input text is cleaned, split into tokens, and embedded, similar to how a language model processes a prompt.

Both paths, along with conditioning signals for loudness, bandwidth, and speaking rate, feed into the Zonos 3 Mini core. This is where the mixture-of-experts routing happens: for each generation step, the model selects a couple of experts out of 64 to predict the next frame of audio codec data. Each predicted frame is fed back into the model to generate the next one, and a decoder converts the stream of frames into 44.1 kHz audio in real time as it’s produced.

Let every team build.
Without shadow IT.

The audio codec itself is worth understanding separately. On the encoding side, audio is compressed down to just 15 frames per second. A dedicated semantic branch, built on a frozen speech model (a W2V-BERT variant), captures the linguistic content of the audio and stores it in a single code book. The remaining acoustic detail, the fine-grained texture of the sound, gets distributed across nine additional code books, for ten codes total per frame. On decoding, these two components (semantic and acoustic) are recombined to reconstruct the final voice output.

This split between “what is being said” and “how it sounds” is a common design pattern in modern neural audio codecs, and it’s part of why MoE-based TTS models can achieve decent voice cloning without needing enormous active compute per token.

Is Zonos 3 Mini worth running locally?

For developers who want multilingual TTS with commercial usage rights and no per-character API fees, Zonos 3 Mini is a reasonable option to self-host. The Apache 2.0 license removes legal friction around commercial deployment, something not every open TTS model offers.

In hands-on testing across multiple languages, including English (US and Irish-accented), Hindi, Telugu, French, Italian, Korean, Spanish, Portuguese, and Chinese, output quality varied. Some languages, like Italian, Portuguese, and the initial American English test, sounded solid. Others, including a test with an Irish accent, came across as less natural, with pronunciation and delivery described as “plasticky” in places. Voice cloning, tested by uploading a personal voice sample and generating Telugu speech from it, preserved reasonable voice texture even when crossing languages, though it wasn’t flawless.

The overall impression from testing is that general TTS quality across the open-source ecosystem may be approaching a plateau, where incremental releases improve coverage (more languages, better licensing, smaller footprints) more than they leapfrog raw output quality. That’s still useful: broader language coverage and permissive licensing matter a lot for real products, even if the audio itself doesn’t sound dramatically better than prior-generation models.

Frequently Asked Questions

What license is Zonos 3 Mini released under?

It’s released under Apache 2.0, which permits commercial use without the restrictions some other open-weight TTS models impose.

How many languages does Zonos 3 Mini support?

It supports 43 languages, covering major world languages as well as regional languages like Telugu and Hindi.

Can Zonos 3 Mini clone voices from a short audio sample?

Yes. The speaker pipeline accepts a reference audio clip and converts it into either a raw audio codec representation or a speaker embedding, which the model uses to imitate that voice during generation, including across different target languages.

How much VRAM do I need to run Zonos 3 Mini?

In testing, the model used close to 25 GB of VRAM on a system with 48 GB available, suggesting a 24 GB GPU should be sufficient, though actual usage can vary with input length and other workloads on the GPU.

Why does a 100-million-active-parameter model need so much VRAM?

Because it’s a mixture-of-experts model, all 64 experts and the full 800 million parameters generally need to be loaded into memory even though only a couple of experts process each token. Combined with the audio codec and embedding components, this pushes total VRAM usage well above what the active parameter count alone would suggest.