Understand this story
On November 9, 2026, Zyphra Research released the ZONOS3-Mini real-time text-to-speech model, which supports zero-shot voice cloning. The model uses a sparse MoE architecture with 100M activated parameters and 800M total parameters, and is open-sourced under the Apache 2.0 license. According to Zyphra Research, the model achieves first-audio latency of under 570 ms on a Pixel 9 in the browser and under 150 ms on an M3 MacBook Air, and provides WebGPU, SGLang, vLLM, and GGML runtimes. The weights are now available on Hugging Face, the runtimes have been released on GitHub, and the ZTTS2-Eval benchmark covering 35 languages has also been launched. There are currently no further developments.
AI-generated from public reports
Latest development · Report
Zyphra Releases ZONOS3-Mini Real-Time Voice Cloning TTS ModelDevelopments
1 confirmed developments
- Source:Zyphra Research
Reports and discussion timeline
Browse reports and discussion by time. Their evidence levels remain distinct.
Nov 9
Zyphra ResearchFirst-partySelected
Zyphra Releases ZONOS3-Mini Real-Time Voice Cloning TTS Model
Zyphra has released ZONOS3-Mini, a real-time TTS model that supports zero-shot voice cloning. It uses a sparse MoE architecture with 100M activated parameters and 800M total parameters, and is open-sourced under the Apache 2.0 license. In the browser, the model achieves first-audio latency of under 570 ms on a Pixel 9 and under 150 ms on an M3 MacBook Air, and it provides WebGPU, SGLang, vLLM, and GGML runtimes. The weights are now available on Hugging Face, the runtimes are released on GitHub, and the ZTTS2-Eval benchmark covering 35 languages has also been launched.