Zyphra Releases ZONOS3-Mini: Open-Source TTS With 100M Active Parameters

Zyphra released ZONOS3-Mini, an open-weights text-to-speech model under Apache 2.0. It is a sparse mixture-of-experts design with 800 million total parameters and about 100 million active per token, and it supports zero-shot voice cloning plus word-level control over pace, emotion and delivery.

The point is where it runs: in a browser via WebGPU, on Apple Silicon, on CPUs through ggml, and on NVIDIA servers. Zyphra reports 80 milliseconds to first audio on a single H100 and under 570 milliseconds in a browser on a Pixel 9. Good speech synthesis is moving from rented cloud service to something that lives on the device.

We're broadly glad to see this. Local, permissively licensed speech models are good for accessibility tools, for apps that shouldn't send audio to a server, and for small developers who can't pay per-character API fees. Running in a browser tab with no install is a meaningful step; it means a screen reader, a language tutor or a game can ship natural voices without a backend.

The engineering story is also a good one. Zyphra says its previous model, ZONOS2, used 8 billion parameters, which limited deployment. This one rebuilt the codec, the backbone and the speaker encoder, and Zyphra says it trained in 2.5 days on 112 H100s. That's a reminder that efficiency, not just scale, is where a lot of useful progress is happening.

Now the caveats, and they're significant. Every benchmark here comes from Zyphra's own ZTTS2-Eval suite. On its Capability Index, ZONOS3-Mini scores 66.2 in speaker-embedding mode, ahead of Fish Audio S2 Pro and MOSS-TTS v1.5. But in reference-clip mode, the way most people would clone a voice from a recording, it scores 44.7. Its perceived-quality score also trails Qwen3-TTS 1.7B. We'd treat the leaderboard claim as a vendor's opening argument until independent testers weigh in.

The harder issue is cloning. Zero-shot voice cloning from a clip of up to 60 seconds, running locally, under a license that allows commercial use, removes most of the friction that has kept voice impersonation somewhat contained. Hosted services can at least log usage, add watermarks or refuse obvious abuse. Weights in a browser can't be recalled. Nothing in the coverage we reviewed describes watermarking or consent checks, and we'd like Zyphra to address that directly.

If this scales, and it will, natural synthetic voices become a commodity layer in apps and devices. That's mostly good. It also means phone scams, fake voicemails and impersonation get cheaper at the same pace. The answer isn't to stop open releases; it's for banks, phone carriers and platforms to stop treating a familiar voice as proof of identity.

Text-to-speech has split between closed hosted services and a growing set of open models. Mixture-of-experts designs activate only a slice of their parameters per token, which is how ZONOS3-Mini keeps compute low. Zyphra also offers a hosted version on its own cloud.