Speech models often trade voice quality and control for the latency and compute needed to run them locally. This release tackles that tradeoff by combining a low-rate audio representation with sparse expert routing: 800 million parameters in total, but about 100 million active per token. The result is designed to make voice cloning and fine-grained delivery control practical beyond a server.
AI Audio2026
ZONOS3-Mini
Clones voices from audio samples and generates steerable speech in 43 languages. Its sparse MoE activates about 100 million parameters per token, enabling real-time use in browsers and on local devices.
Introduction
Information
- Websitehuggingface.co
- OrganizationsZyphra
- AuthorsGabriel Clark, Sofian Mejjoute, George Close, Beren Millidge
- Published date2026/11/09
More Items
AI Model2026
ComfyUI-ready RelayVSR weights upscale video by 2× or 4×. Sparse generative keyframes guide a lightweight network that processes the intervening frames, avoiding the cost of running a large model on every frame.
AI Model2026
Turns images into structured visual data—scene descriptions, objects, text, masks, and spatial relations—on local devices. Its quantized model is about 22 MB, targeting phones, wearables, and robots.
AI Audio2026
Turns English text into speech, with normalization for natural readings of numbers, dates, prices, and codes. The 40.6 MB model runs locally on a CPU with NumPy and includes Juno and Mira voices.