Zyphra has released ZONOS3-Mini, an open-source text-to-speech model meant for real-time use on local devices, MarkTechPost reports. It’s a sparse mixture of experts: 800 million parameters total, about 100 million active per token. Weights are Apache 2.0. It does zero-shot voice cloning and word-level control over pace and emotion, and it can run on NVIDIA GPUs, Apple silicon, CPU, or even in a browser via WebGPU.
Zyphra says it ranks first among open models on the company’s own ZTTS2-Eval Capability Index. Time-to-first-audio on a high-end GPU can be very low; in a phone browser it’s slower. Cloning from a raw reference clip scores much lower than the speaker-embedding mode, and every benchmark here was run by Zyphra.
My take
Small enough to try locally is the selling point. I’ll wait for independent tests before I treat those leaderboard numbers as gospel.
Review before you install: read the release notes yourself, back up your config, and try new versions somewhere you can afford to break first.
God Bless, and Happy hacking!
