ZONOS3-Mini

ZONOS3-Mini is Zyphra's real-time text-to-speech model, announced on 9 November 2026 in San Francisco. The company describes it as a compact zero-shot voice-cloning TTS system that can also run on-device. Officially it is a sparse mixture-of-experts model with about 100 million active parameters and 800 million total parameters, released under Apache 2.0. The post is credited to Gabriel Clark, Sofian Mejjoute, George Close, and Beren Millidge.

Zyphra says training ran on 112 NVIDIA H100 GPUs for 2.5 days. You can use the model on Zyphra Inference (the page says it is served on NVIDIA H200s), on your own GPU, or in a browser on-device. Runtimes shipped at launch include WebGPU in the browser, SGLang and vLLM on servers, and other targets through GGML. The company quotes time-to-first-audio of under 570 ms in a Pixel 9 browser and under 150 ms on an M3 MacBook Air. Those figures come from the launch post; they are not our own measurements.

Highlights:

• Official post: https://www.zyphra.com/our-work/zonos3-mini ;
• Weights: https://huggingface.co/Zyphra/ZONOS3-Mini ;
• Controls cover identity, accent, quality, mood, and delivery, including word-level delivery and a post-generation edit model;
• Relative to ZONOS2, the release emphasizes a smaller model and a new codec so real-time speech no longer depends on a very large parameter count.

Why it matters

Open TTS that can clone a voice, accept fine delivery control, and still run on a phone browser has been hard to offer at once. ZONOS3-Mini is aimed at that gap: Apache 2.0 weights, a WebGPU runtime, plus knobs for accent, apparent age, emotion, pace, and projection, including changes that apply to individual words rather than a whole sentence. For narration, character voices, or an assistant that should stay on the device, that is a concrete alternative to sending every transcript to a cloud API.

The post also introduces Zyphra Audio Codec 1 (ZAC-1), the ZTTS2-Eval test set, and Zyphra's own TTS Capability Index, where the company ranks ZONOS3-Mini first among open models. That ranking is from Zyphra's benchmark. Treat it alongside the listening samples, not as an independently reproduced league table.

Limits

Latency, throughput, and the "first among open models" claim all come from Zyphra's materials; this page does not re-run those tests. Cloning and word-level control will vary with the reference clip, language, and device, and the published examples are largely in English. Voice cloning is a deep-synthesis use in many places, so you need consent and the local labeling rules before shipping it. Pricing, quotas, and cloud SLAs should be taken from Zyphra Inference at the time you use it — do not treat unpublished commercial terms as settled.

1. Read https://www.zyphra.com/our-work/zonos3-mini and listen to the comparison samples on that page first;
2. Get the weights from https://huggingface.co/Zyphra/ZONOS3-Mini and pick the WebGPU browser runtime, SGLang / vLLM, or a GGML runtime as documented;
3. For character work, try utterance-level identity, accent, and emotion controls first, then use word-level delivery or the edit model for local changes;
4. Before cloning a real voice, check consent, privacy, and local deep-synthesis rules; re-measure latency on your target device rather than copying the launch numbers.