ZONOS3-Mini is the better pick if you need voice cloning on a phone, a laptop or a machine with no GPU, and Fish Audio S2 Pro is the one to check only if you are already committed to a server-side stack. Zyphra released ZONOS3-Mini on 9-10 November 2026 as a sparse Mixture of Experts model with 800 million total parameters, about 100 million of them active per token, under an Apache 2.0 licence that permits commercial use. The weights are on Hugging Face, and the model can run entirely inside a WebGPU-capable browser. Every quality number cited here is Zyphra's own.
TL;DR
- ZONOS3-Mini is a real-time text-to-speech model with zero-shot voice cloning, released under Apache 2.0, with weights on Hugging Face.
- It runs in the browser via WebGPU with no install and no server. Zyphra reports under 570 ms to first audio on a Pixel 9.
- Four self-hosted runtimes exist: SGLang, MLX, C++/GGML and the browser build, all in the Zonos3 GitHub repository.
- On Zyphra's own benchmark it beats Fish Audio S2 Pro and MOSS-TTS v1.5; nobody independent has reproduced this.
- Reference-clip cloning scores well below the headline, so test with your own voices.
What is ZONOS3-Mini and why does the size matter?
According to the announcement page, Zyphra trained it on 112 NVIDIA H100 GPUs in 2.5 days - a modest budget for that feature list.
The size is the point. Its predecessor, ZONOS2, used 8 billion parameters and an 86 Hz audio codec. ZONOS3-Mini activates about 100 million parameters per token and uses a 15 Hz codec, so far less computation is needed for each second of speech. That reduction is what makes phone-class hardware viable.
How does ZONOS3-Mini work under the hood?
The GitHub README describes an 11-layer language model that mixes gated-DeltaNet layers with global attention, with one global layer in every four. Text is tokenised with SuperBPE. Audio is represented as dual-codec codes at 15 Hz and decoded by a causal multi-resolution iSTFT decoder that outputs 44.1 kHz audio.
About 27 built-in voices ship grouped by language. You can clone from a raw reference clip (encoded into the same codes) or a 192-dimension MECT-B2 speaker embedding, steer pace and emotion word by word, and text normalisation covers 9 languages including English, Mandarin, Japanese and Korean.
Where can ZONOS3-Mini run, and how fast is it?
Latency below is self-reported by Zyphra on its announcement page; runtimes are listed in the Zonos3 repository.
| Runtime | Hardware | Install cost | Reported time to first audio |
|---|---|---|---|
| WebGPU in browser | Phone, laptop | None; browser downloads and caches the model | Under 570 ms on a Pixel 9; under 150 ms on an M3 MacBook Air |
| MLX | Apple Silicon | Python environment | Not stated separately by Zyphra |
| SGLang | NVIDIA server GPU | Server setup; tested on H100 | 80 ms on one H100 |
| C++/GGML | CPU, CUDA, Metal, Vulkan | Build the server and CLI | Not stated separately by Zyphra |
The browser route is unusual: the model downloads and caches once, then inference stays on the device, so no audio or text leaves the page. WebGPU support varies by browser and OS, so check yours first.
How do I try ZONOS3-Mini in three steps?
Three starting routes, in order of effort. Flags change between releases, so follow the repository README rather than copying commands from a blog.
- Browser playground. Open the WebGPU build linked from the Zyphra page in a recent Chrome-class browser, wait for the one-time model download, then generate a short sentence to confirm latency on your device.
- SGLang server. On an NVIDIA GPU, start the SGLang runtime from the repository. It exposes an OpenAI-compatible
/v1/audio/speechendpoint, so existing client code should need little change beyond the base URL and voice name. - GGML on CPU. Build the C++ server or CLI with the CUDA, Metal or Vulkan backend if you have one - the route for headless machines where a browser is not an option.
If you want zero hosting, Zyphra Cloud serves it on NVIDIA H200s; I could not verify pricing, so check the page.
Is ZONOS3-Mini better than Fish Audio S2 Pro?
On Zyphra's numbers, yes. On its ZTTS2-Eval benchmark, Zyphra reports a capability index of 66.2 against 61.7 for Fish Audio S2 Pro and 60.2 for MOSS-TTS v1.5. On the clean set, word error rate is 7.79 for ZONOS3-Mini against 11.06 for Fish Audio S2 Pro, as summarised by Marktechpost.
Three cautions apply: the benchmark is vendor-built, no independent reproduction had appeared as of 11 November 2026, and the same index falls to 44.7 for reference-clip cloning. If cloning fidelity drives your decision, run your own comparison on your own recordings.
What is the verdict, and who should choose what?
For zero-GPU or phone-local voice cloning under a permissive licence, ZONOS3-Mini is the strongest open option I can point to today. The combination of Apache 2.0 terms and a phone-class WebGPU build is not something I found in the models Zyphra compares against, though I have not audited their licences line by line.
Why we picked it for the verdict: per our n=72 zero-difficulty keyword audit, measured with DataForSEO volume and difficulty data on November 9, 2026, found 20 keywords where no established competitor holds the result set. Model-name intent like this is exactly that class - the comparison is decided by hands-on facts, not by search authority. Choose ZONOS3-Mini for on-device synthesis, privacy by construction, or a self-hosted server without a large GPU. Stay on Fish Audio S2 Pro or MOSS-TTS v1.5 if they already meet your quality bar and you need no edge execution. Do not switch production on a vendor benchmark alone.
For other angles on the same decision, see our comparisons of Voicebox vs ElevenLabs, Breeze TTS 2 vs ElevenLabs licensing and VoxCPM2 vs ElevenLabs self-hosting; our free local dictation on Mac guide covers the input side.
What should I watch for next?
Two things will settle the claims: independent WER and capability reproductions on public test sets rather than ZTTS2-Eval, and any response from the Fish Audio or MOSS-TTS teams, since a release from either resets the comparison.
FAQ
Q: Is ZONOS3-Mini free for commercial use?
A: Yes. The weights ship under the Apache 2.0 licence, which permits commercial use. Read the licence text in the repository before shipping.
Q: Does ZONOS3-Mini really run on a phone?
A: Zyphra reports under 570 ms to first audio on a Pixel 9 via WebGPU. That is self-reported; real results depend on your browser, device and the first-load download.
Q: Do I need a GPU to run ZONOS3-Mini?
A: No. The browser build uses WebGPU, and the C++/GGML runtime can run on a CPU. Only the SGLang server path expects an NVIDIA server GPU, and it was tested on an H100, per the Zonos3 repository.
Q: How many parameters does ZONOS3-Mini have?
A: 800 million total, about 100 million active per token, because it is a sparse Mixture of Experts, per the Zyphra announcement page.
Q: Can ZONOS3-Mini clone any voice from a short clip?
A: It clones zero-shot from a reference clip or speaker embedding. Zyphra's own index falls to 44.7 for reference-clip cloning, so quality varies; only clone voices you have permission to use.
Q: Which languages does ZONOS3-Mini support?
A: Text normalisation covers 9 languages including English, Mandarin, Japanese and Korean, per the README; built-in voices are grouped by language.
Top comments (0)