Zonos 3 Mini: Free Open-Source Voice Cloning, Explained

What Zyphra actually released

Zyphra is a San Francisco speech-model company, and ZONOS3-Mini is its third-generation open model. The headline number is the size: 800 million total parameters, but a sparse mixture-of-experts design means only around 100 million of them run for any given word. That is why it is fast on ordinary hardware instead of needing a data-center GPU.

The model also ships with its own audio codec, ZAC-1, which compresses speech to 44.1 kHz output at just 15 frames per second. Fewer frames per second means less work per second of audio, which is the other half of the speed story. Zyphra trained it on 112 NVIDIA H100 GPUs in 2.5 days, then tuned it with preference optimization to cut down on skipped words and wobbly delivery.

As Zyphra put it in its release announcement: "we release runtimes for browsers via WebGPU, servers via SGLang and vLLM, as well as many others through GGML." In plain terms: the same model is packaged four ways, so it runs on a phone browser, a Mac, a Windows PC, or a server — pick the one that matches your machine.

How fast is it, really?

Zyphra's own figures are striking, and they are worth repeating with the caveat that they are Zyphra's figures, not independent tests. On a single H100 server GPU, the first chunk of audio arrives in a median of 80 milliseconds with eight requests running at once, and peak throughput hits 457 seconds of audio generated per second. Those are server numbers, but the consumer ones matter more: under 570 milliseconds to first audio in a phone browser on a Pixel 9, and under 150 milliseconds on a MacBook Air with an M3 chip.

An independent hands-on test backs up the "runs on a laptop" claim. A tester ran 58 Korean clips on a MacBook Pro with an M4 Pro chip, scored them by transcribing the output back with Whisper, and found the model generating at roughly 10x real-time speed — free and offline, with the core model file weighing just 0.66 GB. The full repository is a hefty 30.7 GB download, but you can skip the cloning encoders (5.6 GB) if you only want the built-in voices.

One honest note: the quality benchmark is Zyphra's own. On its self-built ZTTS2-Eval index, ZONOS3-Mini ranks first among open-source models — but in speaker-embedding mode it scores 66.2, while the simpler reference-clip cloning mode scores only 44.7, and its clean-set quality score of 3.07 trails some rivals. Translation: the best mode is genuinely good; the easiest mode is noticeably weaker. Treat the "number one" claim as a vendor result until independent tests confirm it.

What you can actually do with it

This is where Zonos 3 Mini stands out from most free voices. It is not just a reader — it is a voice designer.

The cloning part deserves a straight warning: cloning a real person's voice needs that person's consent. Our full guide to whether AI voice cloning is legal covers the rules, and the short version is simple — clone your own voice, or a voice you have permission to use, and label AI audio clearly when you share it.

Zonos 3 Mini vs the alternatives you already know

If you just want to hear text read aloud, you do not need Zonos 3 Mini — a free browser tool does that in seconds. The model earns its place when you want something the free tools cannot do: your own voice, offline, with no meter running.

• vs ElevenLabs: ElevenLabs still sounds more polished and needs no setup, but its free tier is capped and cloning sits behind a paid plan. Zonos is free forever and runs locally — you trade polish and convenience for control and zero cost. Our ElevenLabs alternatives roundup covers the middle ground.
• vs Phonon: Phonon is smaller (31M parameters) and simpler, but it has no voice cloning at all. Zonos is the one to pick when cloning is the point. Our Phonon explainer covers that model in full.
• vs Google's Edge Gallery models: Google's bundle (Kokoro-82M, PocketTTS) is aimed at app developers shipping offline readers. Zonos is aimed at anyone who wants a cloneable voice. Same offline idea, different job.

The pattern across all three is the same one we saw last week: tiny, open, on-device voices are having a moment, and the free tier keeps getting better in public, where anyone can try it.

How to try it today (no coding needed)

The honest version: there is no big friendly "Zonos app" yet. But you can hear it in about two minutes.

1. Try Zyphra's browser demo first. The weights are on Hugging Face and the demo runs via WebGPU in a modern browser — no install, no account. Type a sentence, pick a built-in voice, and listen. This is the fastest way to judge the quality with your own ears.
2. If you want the easy path, use a free reader now. TextToSpeechH turns up to 10,000 words into natural speech in 30 voices and downloads the MP3 free, no sign-up. It runs in the cloud, so you need internet — but it is the fastest way to hear a document read aloud today while Zonos apps are still being built.
3. If you are technical, run it locally. The GitHub repository has runtimes for Apple Silicon (MLX), standard CPUs (ggml), servers (SGLang), and the browser (WebGPU). Start with the built-in voices — the full 30.7 GB download is only needed if you want cloning.
4. If you want offline reading without setup, our offline text-to-speech guide lists seven free tools that already work with no connection.

The bottom line

Zonos 3 Mini is the most capable free voice-cloning model you can run yourself right now: real cloning, 43 languages, Apache 2.0, and fast enough to speak in half a second on a phone. It is not yet a finished product for non-technical users, and its headline benchmark is self-reported — but the model itself is free, open, and already running in browsers. If you have ever wanted your own voice to read your words aloud, this is the cheapest way to try it: free, private, and yours to keep.

Try it now: paste any text into our free text-to-speech tool and hear it in 30 voices, or read our AI voice cloning guide to see how cloning works before you try it yourself.