Zyphra has released ZONOS3-Mini, a multilingual text-to-speech model with voice cloning and delivery controls. Its publisher announcement describes 800 million total parameters, about 100 million active per token, and support for 43 languages.
Weights, code and a Space
Zyphra identifies the model weights as Apache 2.0 and its reference code as MIT licensed. An interactive Hugging Face Space is also linked by the publisher. Check the original pages for current availability and terms.
The company reports real-time performance and evaluation results on several devices. Those figures have not been independently verified by us; total memory depends on the weights and runtime, not just the active parameter count. We have not tested the model or demo.
Open weights and local runtimes could make speech workflows more accessible to developers and creators. Anyone testing voice cloning should obtain the speaker's permission and guard against impersonation.
Original reporting based on the linked publisher sources; no independent product or benchmark test.
