Skip to main content

Zyphra ZONOS3-Mini: Open Voice Cloning TTS for Phones and Laptops

ZONOS3-Mini is an open text-to-speech model from Zyphra, released this week with Apache 2.0 weights. It has about 800M parameters but only around 100M active per token, speaks 43 languages, and clones a voice from a short clip. Zyphra reports under 150 ms to first audio on an M3 MacBook Air, though its benchmarks are self-reported.

On this page

What Zyphra released

Zyphra, the AI lab behind the earlier Zonos speech models, published ZONOS3-Mini this week, with a technical write-up dated 11 November. It is a text-to-speech model with zero-shot voice cloning and fine-grained control over how the voice sounds. The weights are licensed under Apache 2.0, and Zyphra's runtimes and reference code are released separately.

The release is broad for a speech model. Zyphra ships PyTorch weights, quantized GGUF files, a Transformers version, a browser playground that runs on WebGPU, and servers for NVIDIA GPUs, Apple Silicon through MLX, and a C++ build that targets CPU, CUDA, Metal, and Vulkan.

A small model built like a big one

ZONOS3-Mini has about 800M parameters in total, but it is a sparse mixture of experts that activates only around 100M per token, choosing 2 of 64 experts at each step. That design is what lets it stay fast on modest hardware while holding more knowledge than a dense 100M model could. The backbone mixes Gated DeltaNet layers with a few full attention layers and can hold about 10 minutes of continuous audio in context.

It covers 43 languages across Europe, the Middle East, and East, Southeast, and South Asia. A new audio codec produces 44.1 kHz output, and voice cloning works either from a reference clip of up to 60 seconds or from a compact speaker embedding that can be saved and reused as a preset.

Speed on local hardware

The headline for anyone who wants private, offline speech is local speed. Zyphra reports time to first audio under 150 ms on an M3 MacBook Air and under 570 ms running in a browser on a Pixel 9. On a single H100 server GPU it reports an 80 ms median with eight requests at once. The model card says it runs in real time on phones and laptops.

If those numbers hold up, a screen reader, a voice assistant, or an offline reading app could generate natural speech on the device without sending text to a cloud service. Zyphra does not publish minimum memory figures, so how well it runs on older or cheaper phones is still unclear.

Control and the cloning question

Beyond basic speech, Zyphra exposes controls for speaking rate, loudness, age, accent, emotion, and whispering, and even lets users steer delivery for specific words while audio streams. A speech editing checkpoint can replace a word or phrase inside an existing recording.

That capability cuts both ways. A small, open, high quality cloning model on a laptop is also a tool for impersonation, and the model card does not set out any usage restrictions on cloning voices. Anyone deploying it should get consent from the people whose voices are cloned.

How much to trust the claims

Every number above comes from Zyphra. It ranks ZONOS3-Mini first among open-weight models on its own Capability Index, ahead of Fish S2 Pro and MOSS TTS v1.5, using a benchmark it designed. Zyphra also says some controls are approximate: its England accent approximates Southern English, age settings are targets rather than measured ages, and emotion steering can shift pitch and loudness. Independent reviewers have flagged that the rankings are not yet verified.

Even with those caveats, a permissively licensed model of this size with phone and browser runtimes from day one makes local text-to-speech much easier to build.

CuriousLM runs supported AI models locally on your device. Try CuriousLM.