Original Title: Zyphra/ZONOS3-Mini
Zyphra's open-source text-to-speech model has a total of 8 billion parameters, with approximately 1 billion activated per token in its MoE architecture, supporting 43 languages (including Mandarin, Japanese, and Korean). It can clone a voice using a reference audio clip (up to 60 seconds long), or use a speaker vector to adjust character timbre, with control over loudness, bandwidth, and speech rate via tiered settings; the official team claims it can also control emotion and rhythm at the word level. It outputs 44.1 kHz audio, with a maximum generation length of 10 minutes per request; officially, on a single H100 GPU with 8 concurrent streams, the median first-packet audio latency is 80 milliseconds, and peak throughput reaches 457 seconds of audio generated per second.
Why it matters Clone voices in the browser using local GPU, control emotion and speech rate at the word level, and use Apache-licensed weights for commercial purposes
Upvotes 83 · Likes 84 · Downloads 856
Reasons
- Small footprint, officially claimed to run in real-time on phones, laptops, and browsers; the browser version uses WebGPU, requiring no installation or server
- Multiple deployment options: SGLang-Omni provides an OpenAI-compatible /v1/audio/speech interface with streaming output, while official runtimes are available for CUDA, MLX (Apple Silicon), C++/ggml, and WebGPU, including 27 default voices
- Weights are Apache-2.0 licensed, allowing commercial use
Usage
- Try it directly in the browser via the official Zonos Voice Studio or Lite page (requires a WebGPU-supported browser; the model will be downloaded on first access)
- Server-side: After installing SGLang-Omni, run
sgl-omni serve --model-path Zyphra/ZONOS3-Mini; addreferencesin the request to specify reference audio for cloning, and add"stream": truefor streaming output
Risks
- Quality and speed comparisons were measured by the official team on their own ZTTS2-Eval benchmark; full evaluations are available in the official blog
- Voice cloning can be misused; obtain authorization before using someone else's voice
- This entry has not been tested; refer to the model card for actual performance