✦
Summary
ZONOS3-Mini is Zyphra's compact zero-shot voice-cloning text-to-speech model, released in November 2026. It is a sparse mixture-of-experts model with 100M active and 800M total parameters, trained in 2.5 days on 112 H100 GPUs, and it runs in real time on-device in a browser, on a phone, a laptop or a server, with under 570 ms time-to-first-audio in the browser on a Pixel 9 and under 150 ms on an M3 MacBook Air. It is a new model generation and a distinct record from the earlier Zonos TTS in the directory: a successor to the 8B ZONOS2 with a new codec, ZAC-1, and new controllability.
Key Features
Zero-shot voice cloning from a reference clip with no fine-tuning
Real-time on-device synthesis: under 570 ms time-to-first-audio in the browser on Pixel 9, under 150 ms on M3 MacBook Air
Fine-grained voice control over identity, accent, quality, mood and delivery, including word-level delivery steering
Voice-design controls for age, gender, accent and texture such as breathy or raspy
Delivery steering for whisper, emotion, pace, projection and liveliness, per word or per span
Sentence-edit model for seamless edits with the same controls
Runtimes for WebGPU in the browser, SGLang, vLLM, MLX and C++ with GGML
43 languages with 27 default voices, ranking first among open models on ZTTS2-Eval
About ZONOS3-Mini
The model clones a voice from a reference clip with no fine-tuning, and offers fine-grained control over identity, accent, quality, mood and delivery, including word-level delivery steering for whisper, emotion, pace, projection and liveliness. Voice-design controls cover age, gender, accent and texture such as breathy or raspy, and a sentence-edit model makes seamless edits with the same controls. It speaks 43 languages with 27 default voices, and ranks first among open models on the ZTTS2-Eval benchmark.
ZONOS3-Mini is Apache 2.0 open weights on Hugging Face as Zyphra/ZONOS3-Mini, with MIT-licensed runtimes, so the model itself is free with no paid tier. Users can try it in the browser WebGPU playground or a Hugging Face Space, download the weights, or self-host with SGLang, vLLM, MLX or C++ and GGML runtimes. Zyphra also offers cloud inference hosting for the model separately.