Zyphra Releases ZONOS3-Mini Open-Source Text-to-Speech Model
Zyphra has launched ZONOS3-Mini, a compact 100M-parameter open-source text-to-speech and voice cloning model supporting 43 languages. It operates in real time on consumer hardware and directly in web browsers via WebGPU. This release brings high-fidelity voice cloning, emotion steering, and multilingual support to edge devices and web applications without relying on heavy cloud infrastructure. Releasing the model weights under the Apache 2.0 license empowers developers and researchers to easily integrate advanced audio AI locally. ZONOS3-Mini features an 11-layer architecture with 100M active parameters per token using a latent Mixture of Experts design and Gated DeltaNet layers. It supports native runtimes for CUDA, Apple Silicon, C++/ggml, and WebGPU, generating 44.1 kHz audio.
## BACKGROUND
Text-to-speech (TTS) models synthesize spoken audio from written text, while modern voice cloning allows AI to replicate a specific person's voice from a short audio sample. Running these models locally on consumer hardware or browsers has traditionally been difficult due to high computational demands, making lightweight and efficient architectures like ZONOS3-Mini notable for developers.