Zonos2 (ZONOS2)
modelYour notes
Zyphra's open-weight multilingual text-to-speech model and successor to Zonos — an MoE backbone (16 experts, top-1 routing; 28 layers, GQA) generating DAC audio tokens from UTF-8 byte input, with ECAPA-TDNN speaker embeddings for zero-shot voice cloning. Trained on 6M+ hours of multilingual speech; Apache-2.0, with a Mini-SGLang-based inference server for low-latency synthesis.
Three-tier language coverage: Tier 1 English / Mandarin / Japanese, Tier 2 adds Korean, Russian, Italian, Portuguese, French, Spanish, Vietnamese, German, Hebrew, Dutch, and Tier 3 a further ~23 languages. Extends Zyphra's foundation-model line beyond language and EEG into speech generation.
ZONOS3-Mini (2026-11-09, Apache 2.0) is the next generation in a smaller size. It is a sparse MoE with 800M total and 100M active parameters, and Zyphra says it was trained on 112 H100s in 2.5 days. The backbone moves to Zyphra's language-model components:
- Mixing and residuals: a 3:1 hybrid of Gated DeltaNet and global attention over 11 layers, a latent MoE, and attention residuals.
- Input and codec: it reads SuperBPE text tokens and predicts codes from Zyphra's new ZAC-1 dual codec at 15 Hz, replacing 86 Hz DAC. Zyphra reports a tenfold drop in word error rate over a DAC-trained equivalent.
- Post-training: DPO followed by GRPO.
On Zyphra's own TTS Capability Index it ranks first among open models, ahead of Fish S2 Pro and MOSS TTS v1.5. Zyphra gives no numeric scores in the text, and it has not announced a larger ZONOS3.