Playbooks · Everyday tasks
ZONOS3-Mini: open, multilingual, real-time voice
Written November 12, 2026
A voice model you can run in a browser, on a phone, or on your own server, in 43 languages, with the weights under Apache 2.0. Zyphra's ZONOS3-Mini, posted November 11, is built to start speaking in well under a second.
What was released
Zyphra's ZONOS3-Mini post announces an open text-to-speech model for multilingual synthesis and zero-shot voice cloning across 43 languages. The model is a sparse mixture-of-experts design with 100 million active parameters out of 800 million total, using a new 15 Hz audio codec, ZAC-1, in place of the 86 Hz codec its predecessor modelled. The weights are Apache 2.0 on Hugging Face, including quantized versions, and the runtimes and reference code are MIT licensed, with serving through SGLang-Omni, native Apple Silicon support, a standalone C++ runtime, and an in-browser WebGPU version. There is a Hugging Face Space to try cloning and synthesis without installing anything. Zyphra says the model was trained in 2.5 days on 112 NVIDIA H100 GPUs.
Real time, in Zyphra's numbers
The latency figures are Zyphra's, stated as its measurements. Time to first audio is under 570 milliseconds in a browser on a Pixel 9, under 150 milliseconds on an M3 MacBook Air, and 80 milliseconds at the median on a single H100, where Zyphra also claims 457 times real-time peak throughput. Those three numbers describe three very different machines, and only the first two describe hardware most readers own. Zyphra also released its evaluation, ZTTS2-Eval, covering 35 clean and 20 in-the-wild language subsets, and says ZONOS3-Mini ranks first among open-weight models on its capability index. That ranking is the maker's own benchmark and index, and this page reports it as such.
Control, not just speech
Two speaker modes are offered. A reference clip of up to 60 seconds can be placed in context for cloning, or a compact 192-dimension speaker embedding can be used for presets and continuous identity steering. The post also presents the model as a testbed for mechanistic voice and delivery control: activation steering that adjusts how a voice sounds and delivers lines without retraining. Post-training with preference optimization and reinforcement learning is described as reducing hallucinations and improving pronunciation stability and multilingual robustness, in Zyphra's account.
What mini means
The comparison that frames the release is with Zyphra's own ZONOS2, which relied on an 8B-parameter backbone and, in the post's words, constrained local and client-side deployment. Mini is the answer to that constraint: a far smaller active footprint that fits browsers and phones. The trade a small model usually makes, in voice range and fidelity against the largest systems, is one to judge from the post's side-by-side samples against open and proprietary systems rather than from its parameter count. What is not in question is the shape of the offer: open weights, open runtimes, and an evaluation anyone can rerun.
Limits
- Every quality and speed figure here is Zyphra's. No independent measurement is cited in the release post.
- Voice cloning of real people carries consent questions the technology does not settle. The post presents cloning as a feature; whose voice you clone is your responsibility.
- 43 languages is the synthesis claim. The released evaluation covers fewer subsets, 35 clean and 20 in the wild, so coverage and measured quality are not the same list.
Related playbooks
- Gemini's Nov 17 shutdowns: Gems and voices
- Pack for the forecast, checked again at pack time
- Negotiate a real job offer without bluffing: your walk-away line, three asks and the words to send
- Set a freelance rate that survives fees, bidding costs, and review time
Sources
Back to: Everyday playbooks hub · Also: Gemini's Nov 17 shutdowns: Gems and voices