Zyphra releases compact ZONOS3-Mini speech model for local devices
Zyphra releases compact ZONOS3-Mini speech model for local devices
Event date: 2026-11-09; precise time unverified.
Zyphra announced ZONOS3-Mini on November 9, releasing a compact text-to-speech model aimed at running on local devices as well as servers. This is catch-up coverage of that release, following a November 10 report from MarkTechPost, rather than a new November 11 launch.
The model uses a mixture-of-experts architecture with about 800 million total parameters and 100 million active parameters. Its official Hugging Face card identifies the weights as Apache 2.0 and describes voice cloning, multilingual speech and controls over how individual words are delivered. The repository’s file listing includes native and browser-oriented model directories.
Local deployment is the main practical distinction. Zyphra provides a WebGPU browser path, Apple Silicon support through MLX, and a C++/ggml runtime for several hardware backends. Its model card also describes streaming speech through an OpenAI-compatible endpoint using SGLang-Omni. These routes give developers options beyond sending every synthesis request to a hosted service.
Zyphra reports browser time to first audio below 570 milliseconds on a Pixel 9 and below 150 milliseconds on an M3 MacBook Air. Those figures are company measurements on specific devices, not guarantees for other phones, browsers or workloads. This review did not run the model or independently measure its audio quality or latency.
MarkTechPost’s coverage highlights a further qualification: the model’s strongest reported benchmark result uses speaker embeddings, while reference-clip mode scores substantially lower. The evaluation was built and run by Zyphra. Developers comparing speech systems should therefore test the conditioning mode they actually intend to use, rather than assuming a headline benchmark score applies equally to every workflow.
The release expands the available options for locally hosted narration and voice interfaces. Its deployment flexibility is verifiable in the documentation; broader claims of superior speech quality still require independent testing.
Sources
- https://www.zyphra.com/our-work/zonos3-mini
- https://huggingface.co/Zyphra/ZONOS3-Mini
- https://huggingface.co/Zyphra/ZONOS3-Mini/tree/main
- https://github.com/Zyphra/ZONOS3-Mini
- https://www.marktechpost.com/2026/11/10/zyphra-zonos3-mini-open-source-tts/
Sources and citations
Revision history
Published .