ZONOS3-Mini

ZONOS3-Mini is Zyphra’s model for generating speech from text. Optional reference audio guides the voice through zero-shot voice cloning. Zyphra’s announcement is dated 9 November 2026; its later Hugging Face publication is a separate publication date.

Text and a reference voice guide the output

The model card lists 43 languages, a 32,768-position context and support for up to 10 minutes of generated audio. Reference voice audio can be up to 60 seconds long. The reference recording and requested output occupy different roles in the generation process.

Use a voice only with the speaker’s permission or another valid right to do so. Clear labelling is useful when listeners could mistake generated speech for a recording of a real event.

Open weights and several runtime options

The model has an 800-million-parameter backbone, with 100 million parameters active. Its weights use Apache 2.0, while the published browser and native runtimes use MIT licences. Zyphra also offers access through its cloud service.

The availability of browser execution does not imply the same speed on every device. Generation time depends on hardware, runtime, text length and output settings. Test the intended device before treating a developer demonstration as a production latency target.

Listen to the details that affect your use

Useful checks include pronunciation of names and numbers, the handling of punctuation, pauses and consistency across longer passages. A voice that sounds convincing in a short sample can behave differently over several minutes.

The developer documentation establishes the model’s published limits and licensing. The sources checked for this page do not establish a current hosted API rate. Check Zyphra Cloud’s terms separately when estimating the cost of a hosted deployment.

Image: Zyphra’s official launch artwork.