Zyphra unveils ZONOS3-Mini: a compact, open-source tts model with 100m active parameters

Zyphra has introduced ZONOS3-Mini, an open-source text-to-speech model designed for real-time operation on local hardware, such as laptops or browser tabs. This sparse Mixture of Experts (MoE) model features 800M total parameters but only 100M active parameters, supporting zero-shot voice cloning and word-level delivery control. The model's weights are available under the Apache 2.0 license. Developed by training on 112 NVIDIA H100s over 2.5 days, ZONOS3-Mini builds upon the earlier ZONOS2 model. The new version incorporates advanced controls, including speaker embedding vectors for identity (age, gender, accent, texture) and activation steering for delivery (pace, emotion, projection, liveliness). Furthermore, a separate edit model, based on a VoiceCraft recipe, allows for post-generation sentence modification. Evaluation on Zyphra's ZTTS2-Eval benchmark, which covers 35 languages of clean speech and 20 in-the-wild languages, shows ZONOS3-Mini achieving a Capability Index score of 66.2, surpassing Fish Audio S2 Pro (61.7) and MOSS-TTS v1.5 (60.2). While the model shows a 7.79 WER on the clean set, Zyphra noted caveats, including a lower MSR-UTMOS quality score of 3.07 compared to Fish S2 Pro (3.16) and Qwen3-TTS 1.7B (3.54).