Zyphra open-sources ZONOS3-Mini MoE TTS at 80ms first audio

Zyphra's new text-to-speech model is 100 million active parameters, trained in two and a half days on 112 H100 GPUs, and released under Apache 2.0. Its central claim lives in the audio codec.

"Under matched data constraints, a TTS model trained on ZAC-1 achieves a 10× reduction in Word Error Rate (WER) compared to one trained on DAC while simultaneously improving acoustic similarity," Zyphra writes in a Hugging Face release post. ZAC-1 runs at a 15 Hz frame rate with 44.1 kHz output, combining one Wav2Vec2-BERT semantic codebook with nine RVQ acoustic codebooks.

The latency numbers are the other hook. Zyphra reports 80 ms median time-to-first-audio on a single H100, under 150 ms on an M3 MacBook Air, and under 570 ms in a Pixel 9 browser. Peak throughput on H100 is given as 457 times real-time. The model handles 43 languages across European, Middle Eastern, East Asian, Southeast Asian, and South Asian scripts.

Architecturally it is a 100M-active, 800M-total sparse mixture-of-experts: 64 experts with top-2 routing from the first layer, nine Gated DeltaNet recurrent layers paired with two global GQA attention layers, and a 32,768-position context window the authors put at up to ten minutes of continuous audio.

On the Zyphra-authored ZTTS2-Eval benchmark, which spans 55 language subsets and 420 items, "ZONOS3-Mini ranks first among open-weight models, ahead of Fish S2 Pro and MOSS TTS v1.5." The leaderboard is published by the same team that trained the model.

Dwarkesh is quoted in the release post: "it no longer has that synthetic, slop-y cadence you expect from these models. It actually feels like you're talking to a person, right?" It lands in a run of open-source releases our tracker has been counting through the fall.