ZONOS3-Mini is Zyphra's open-source text-to-speech model, officially reported to have approximately 1 hundred million active parameters, capable of running in real time on smartphones, laptops, and browsers. It supports voice cloning from a single audio clip, or designing character voices based on age, accent, and audio quality, with word-level control over emotion and speech rate; text normalization covers languages including English, Chinese, Japanese, Korean, German, Spanish, French, Italian, and Portuguese. The weights are released under Apache 2.0, with additional GGUF and WebGPU shards provided.
Why it matters The Apache license allows commercial use. Ready-to-run implementations are available for browser WebGPU, MLX, ggml, and SGLang, lowering the barrier to entry for dubbing, narration, and voice assistants.
Hugging Face likes 13 · Hugging Face trending score 13 · Hugging Face downloads 0
License: Commercial use allowed · ComfyUI: Not available · VRAM: ≈1 GB estimated
Reasons
- Hugging Face text-to-speech trending list
- Score 52 (fresh 15 / source 5 / heat 4 / comfy 0 / vram 13 / demo 5 / license 10)
Usage
- Browser: Open the official Zonos Voice Studio or Lite page to run locally using WebGPU; the model will be downloaded and cached on first access
- Server-side: Launch using sgl-omni serve from SGLang-Omni, utilizing the OpenAI-compatible /v1/audio/speech endpoint; for Apple Silicon and CPU, use the official repository's MLX and ggml runtimes
Risks
- Performance and evaluations are only from the official blog and have not been independently verified; the model card does not separately specify Chinese pronunciation quality
- Voice cloning poses a risk of impersonating others' voices