Zyphra launches open-source ZONOS3-Mini speech model
Zyphra has released ZONOS3-Mini, an open-source text-to-speech model designed for real-time operation on local devices. Its sparse Mixture of Experts architecture contains 800M parameters, with about 100M active for each token through 64 experts and top-2 routing. The weights are available under the Apache 2.0 license, and the model supports zero-shot voice cloning and word-level control of delivery.
ZONOS3-Mini can run in a browser through WebGPU, on Apple Silicon with MLX, and on CPU, CUDA, Metal or Vulkan through ggml. It supports NVIDIA GPUs through SGLang-Omni and can generate up to 10 minutes of audio in one request. Zyphra says the model was trained on 112 NVIDIA H100s over 2.5 days.
In Zyphra’s self-reported ZTTS2-Eval results, the model ranked first among open models on its Capability Index. Reported performance includes 80 ms median time to first audio on one H100 with eight concurrent requests and under 570 ms in a browser on a Pixel 9. Zyphra notes that the benchmarks are internally run and that reference-clip cloning scores substantially lower than speaker-embedding mode.
Written by Sakhanda Wire AI from reporting by MarkTechPost. Facts, quotes and full context are in the original article below.