ZONOS3-Mini: Zyphra Voice Cloning TTS Setup and Pricing

Last updated: November 12, 2026

ZONOS3-Mini is Zyphra’s new open-weight text-to-speech model, released November 9, 2026. It clones a voice from a short sample, speaks 43 languages, and lets you steer accent, age, emotion and pace. With about 100 million active parameters, it runs in real time on phones, laptops and browsers, under the Apache 2.0 license.

Contents

What is new
Specs, license and pricing
How to run it
Voice and delivery controls
How it compares
Limitations
Frequently asked questions

This guide draws on Zyphra’s release post, the ZONOS3-Mini model page, the Hugging Face model card and Zyphra’s Hugging Face blog post.

What is new

Zyphra calls ZONOS3-Mini its “next-generation real-time Text-To-Speech (TTS) model with high-fidelity voice cloning.” It is a sparse mixture-of-experts model with 100 million active and 800 million total parameters, trained on 112 NVIDIA H100 GPUs in 2.5 days.

The main change is efficiency. Zyphra trained a new audio codec, Zyphra Audio Codec 1 (ZAC-1), that runs at 15 Hz instead of the 86 Hz of the Descript Audio Codec. Zyphra says a model trained on ZAC-1 cut word error rate tenfold compared with one trained on DAC under similar data constraints.

The backbone borrows recent LLM designs: a 3:1 mix of Gated DeltaNet and global attention layers, a latent mixture of experts with 64 experts and top-2 routing, and attention residuals. Post-training uses DPO followed by GRPO to reduce hallucinations and improve pronunciation.

Zyphra reports under 570 ms time to first audio in the browser on a Pixel 9, under 150 ms on an M3 MacBook Air, and an 80 ms median on one H100 at a concurrency of 8.

Specs, license and pricing

Item	Official value
Release date	November 9, 2026
Parameters	About 100M active; 800M total in the release post (the model page lists 886M)
Context	32,768 positions, up to 10 minutes of audio per generation
Output audio	44.1 kHz, streamed frame by frame
Languages	43
Voice cloning	Reference clip up to 60 s, or a 192-d MECT-B2 speaker embedding
Default voices	27, grouped by language, in `gguf/voices/`
License	Weights Apache 2.0; runtimes and reference code MIT
Hosted API price	$10.00 per 1M UTF-8 bytes of input text
Hosted model name	`zyphra/zonos3-mini`

Values come from the model page and the model card. Zyphra prices the hosted API by input text bytes, not by audio seconds.

How to run it

There are four official routes: the browser, SGLang-Omni on NVIDIA GPUs, Zyphra’s native runtimes, and the hosted API. To serve it on your own GPU:

Install SGLang-Omni by following its installation guide.
Start a server with sgl-omni serve --model-path Zyphra/ZONOS3-Mini --port 8000.
Wait until curl http://localhost:8000/health reports "healthy".
Send text to the OpenAI-compatible /v1/audio/speech endpoint to get a WAV file.
To clone a voice, add "references": [{"audio_path": "..."}] to the request. To stream 44.1 kHz audio as it is generated, add "stream": true, "response_format": "pcm".

sgl-omni serve --model-path Zyphra/ZONOS3-Mini --port 8000

curl -X POST http://localhost:8000/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"input": "Hello from ZONOS3-Mini!"}' \
  --output output.wav

For laptops and edge devices, the MIT-licensed GitHub repository includes a CUDA server, an MLX server for Apple Silicon, a C++/ggml server and CLI (CPU, CUDA, Metal or Vulkan), and a WebGPU build. They load the GGUF files in the model’s gguf/ folder. For background on these engines, see our SGLang 0.5.23 guide and llama.cpp 0.9.0 guide.

To use the hosted version, Zyphra’s model page shows the standard OpenAI SDK pointed at its own base URL:

from openai import OpenAI
import os

client = OpenAI(api_key=os.getenv("ZYPHRA_API_KEY"),
                base_url="https://api.zyphracloud.com/api/v1")
response = client.audio.speech.create(
    model="zyphra/zonos3-mini", voice="Savannah", input="Hello!")
with open("speech.wav", "wb") as f:
    f.write(response.read())

Voice and delivery controls

Control is the main feature for power users. Zyphra splits it into identity and delivery.

Control	What it changes	How it works
Identity	Age, gender presentation, softness, raspiness, brightness, depth	Directions in the speaker-embedding space
Accent	For example British or Australian	Dedicated task vectors and adapters
Delivery	Pace, emotion, projection, liveliness	Activation steering at inference time
Word-level delivery	Whisper or stress on a chosen span	Alignment attention heads during streaming
Speech editing	Replace a word or phrase after generation	Edit model, `zonos3-ptq-edit.gguf`
Acoustic buckets	Loudness, bandwidth, speaking rate	Bucketed conditioning inputs

Zyphra’s samples use values such as age: +1.5, british: 1.5 and pace: +2. It notes that these are targets: emotion settings “do not guarantee a specific emotion,” and pace steering can also affect loudness and pitch.

A practical workflow for games or narration is to clone a base voice once, save its speaker embedding as a preset, and then vary age, accent and delivery per character rather than recording new reference clips.

How it compares

Against its own predecessor, the change is size. Zyphra says ZONOS2 needed an 8B-parameter backbone on an 86 Hz codec, which limited deployment. ZONOS3-Mini moves to a 15 Hz codec and about 100M active parameters.

Against other open models, Zyphra built a new benchmark, ZTTS2-Eval, covering 35 clean and 20 in-the-wild language subsets. On its Zyphra TTS Capability Index, Zyphra says ZONOS3-Mini “ranks first among open models followed by Fish S2 Pro and MOSS TTS v1.5.” These are Zyphra’s own results, so test on your own voices and languages.

If you want a hosted speech API from a larger vendor instead, compare our guides to Voxtral 3, Qwen-Audio 3.1 and Gemini Voice Remixing.

Limitations

Text normalization is built in for nine languages (English, Mandarin, Japanese, Korean, German, Spanish, French, Italian and Portuguese). In the other languages, text passes through as written, so spell out numbers and abbreviations yourself.
The browser studio needs a browser with WebGPU support and downloads the model on the first visit.
SGLang-Omni serving targets NVIDIA GPUs. Use the MLX or ggml runtimes on Apple Silicon or CPU.
Steering values are generation targets. Zyphra advises judging perceived traits and speaker consistency by listening.
Benchmark and latency figures are Zyphra’s own and have not been independently verified.
Voice cloning can imitate real people. Get consent before you clone someone’s voice.

Frequently asked questions

What is ZONOS3-Mini?

ZONOS3-Mini is an open-weight text-to-speech model from Zyphra, released November 9, 2026. It clones a voice from a short audio sample, speaks 43 languages and lets you steer accent, age, emotion and pace. It is a mixture-of-experts model with about 100 million active parameters.

Is ZONOS3-Mini free?

Yes, the weights are free to download from Hugging Face under the Apache 2.0 license, and Zyphra’s runtimes are MIT-licensed. If you would rather not host it, Zyphra’s hosted API lists a price of $10.00 per 1 million UTF-8 bytes of input text.

Can ZONOS3-Mini run in a browser?

Yes. Zyphra’s Zonos Voice Studio runs the model on your own GPU through WebGPU, with no install and no server. The first visit downloads the model and the browser caches it. Zyphra reports under 570 ms to first audio on a Pixel 9 in the browser.

How long a reference clip does ZONOS3-Mini need for voice cloning?

Zyphra’s documentation says the in-context reference mode accepts a clip of up to 60 seconds, and SGLang-Omni cloning needs no transcript of the clip. A second mode uses a single 192-dimension MECT-B2 speaker embedding for compact voice presets.

How does ZONOS3-Mini compare with ZONOS2?

ZONOS2 used an 8B-parameter backbone on an 86 Hz audio codec, which limited where it could run. ZONOS3-Mini uses a new 15 Hz codec, ZAC-1, and about 100M active parameters, so it can run in real time on phones, laptops and browsers.

ZONOS3-Mini: Power-User Guide to Voice Cloning, Setup and Pricing