ZONOS3-Mini (often written ZONOS3 or Zonos 3) is Zyphra's open-source text-to-speech model that clones a voice from a short recording and speaks 43 languages. Zyphra released it on 9 November 2026 with the weights under Apache 2.0, so what you make can be used commercially. It is small enough to run on your own PC, and in Promptus you can run it with one click from a CosyTemplate, locally or online. Watch the full video test, where we cloned a real voice and scored ZONOS3 against ElevenLabs in eight rounds.
| Spec | ZONOS3-Mini |
|---|---|
| Developer | Zyphra |
| Release | 9 November 2026 |
| Parameters | About 800 million in total, about 100 million active (sparse mixture of experts) |
| Languages | 43 |
| Audio | 44.1 kHz output |
| Voices | 27 built-in voices, plus zero-shot cloning from a short clip |
| Licence | Weights Apache 2.0, inference code MIT |
| Where it runs | Windows, Mac, Linux, browser (WebGPU) and phones, per Zyphra |
ZONOS3-Mini is a small language model for speech. It reads your text plus a description of the voice (a cloned speaker, a built-in voice or a designed one) and predicts audio tokens, which Zyphra's own audio codec, ZAC-1, turns back into a 44.1 kHz waveform. Because it is a mixture of experts, only about 100 million of its 800 million parameters work on each step, which is why it is fast enough for laptops and phones. Zyphra reports first audio in under 150 ms on an M3 MacBook Air and under 570 ms in a browser on a Pixel 9.
Give it 10 to 30 seconds of one person speaking and it copies that voice, with no training step. In the ComfyUI-ZONOS3 node there are two clone methods: voice profile turns the clip into a short fingerprint and works with every slider, and raw clip uses the whole recording.
You can also make a new voice without any recording, with 25 sliders (gender, age, pitch, room, pace, happy, calm and more), 15 accents and 8 presets such as a chatty podcaster or a late-night radio host.
We cloned one real voice with ZONOS3-Mini and with an ElevenLabs v4 voice clone (default settings), ran the same eight lines through both, one take each, and scored how closely each sounded like the real speaker with two speaker-matching models. The real voice scores 91, which is the ceiling.
| Round | ZONOS3 | ElevenLabs | Point |
|---|---|---|---|
| Plain | 79 | 69 | ZONOS3 |
| Excited | 65 | 57 | ZONOS3 |
| Whisper | – | – | ElevenLabs (it really whispered) |
| Sad | 75 | 69 | ZONOS3 |
| Angry | 67 | 70 | ElevenLabs |
| Tongue twister | 81 | 73 | ZONOS3 |
| Long (about 60 s) | 88 | 77 | ZONOS3 |
| Spanish | 70 | 60 | ZONOS3 |
ZONOS3 won 6 rounds to 2 on sounding like the real voice. ElevenLabs still sounds more polished and is better at acting, especially whispering and anger. On an NVIDIA RTX 5090 Laptop GPU, ZONOS3 made six to seven seconds of speech in 0.67 to 0.83 seconds.
Prefer ComfyUI? The free ComfyUI-ZONOS3 node pack adds the same nodes. Our step-by-step ZONOS3 guide covers both routes, with settings and fixes.
Sources: Zyphra's ZONOS3-Mini announcement and the ZONOS3-Mini model card on Hugging Face. Speed and head-to-head results were measured by Promptus.
Only clone your own voice, or a voice you have permission to use.
Join our distributed GPU compute network. Help us make AI accessible, scalable
and secure for designer, developers and start-ups.