ZONOS3-Mini: Zyphra's compact text-to-speech model with voice cloning under Apache 2.0

On November 9, 2026, the model Zyphra/ZONOS3-Mini was published on HuggingFace — a text-to-speech generator with voice cloning and control over emotion, pace and delivery down to individual words. The model supports 43 languages, including Russian, and, according to the authors, runs in real time on phones and laptops, including directly in the browser. The model weighs 28.5 GB in the repository, the largest file being autoencoder/generator.pt at 7.4 GB. The weights are published in pt, safetensors, gguf, pth and bin formats. The architecture is autoregressive, launched via zonos3. The license is Apache 2.0, meaning the weights can be used in commercial projects without revenue restrictions.

What it means

In our audio generators section, text-to-speech is represented by two open source tools, and ZONOS3-Mini differs from both in its terms of use. XTTS from Coqui is also text-to-speech with voice cloning, but its weights are distributed under the Coqui Public Model License 1.0.0: non-commercial use only, with direct or indirect payment from using the model prohibited. ZONOS3-Mini under Apache 2.0 has no such restrictions. At the same time, XTTS is noticeably lighter: the largest weight file is 1.7 GB versus 7.4 GB for ZONOS3-Mini. Fish Speech from Fish Audio is another text-to-speech generator. Its code permits commercial use only with annual revenue up to 500,000 dollars, and the weights are under CC BY-NC-SA 4.0, a non-commercial license requiring the terms to be preserved upon distribution. ZONOS3-Mini under Apache 2.0 is freer in both code and weights. In size, Fish Speech is lighter: 1.2 GB largest file versus 7.4 GB. YuE is a music generator, not a speech one, so it does not directly overlap with ZONOS3-Mini, but for comparison: its weights are also under Apache 2.0, the largest file is 4.6 GB, and the author specifies 24 GB of VRAM to run it. ZONOS3-Mini, as stated by the developer, runs on phones and laptops — the author does not specify VRAM requirements. ZONOS3-Mini is the first text-to-speech model in our reference with weights under Apache 2.0 without restrictions on commercial use and with claimed operation on mobile devices. In size it is heavier than existing analogues, but the license makes it suitable for projects where XTTS and Fish Speech do not fit.