Zonos3-Mini speaks 43 languages from 100M active parameters, free under Apache 2.0

Zyphra has published ZONOS3-Mini on Hugging Face, an open text-to-speech model with voice cloning for 43 languages. It has 800M total parameters but about 100M active per token, and its weights carry an Apache 2.0 license.

Small active size, long audio, many languages

According to the model card, ZONOS3-Mini is a mixture-of-experts model: 64 experts with top-2 routing inside an 11-layer backbone with a model dimension of 1,024. The card lists 800M total parameters, while the page metadata shows 0.9B. The context is 32,768 positions, and a single generation can produce up to 10 minutes of audio.

Speech is handled as tokens. The card describes a codec running at 15 Hz and 44.1 kHz, with 10 codebooks of 4,096 entries each, one semantic and nine acoustic. Text goes through an 8,192-entry tokenizer with NeMo text normalization. Zyphra says the model covers 43 languages, including English, Mandarin, Hindi, Arabic and Vietnamese, and ships 27 default voices grouped by language.

To clone a voice, users supply a reference clip of up to 60 seconds, or a single 192-dimensional speaker embedding. The Python example on the card builds that embedding from a clip of 5 to 10 seconds. Loudness, bandwidth and speaking rate can be controlled as well.

Where it runs, from a browser tab to a GPU server

The deployment list is unusually wide for a speech model. The card says the model runs in the browser on WebGPU with no install, through apps called Zonos Voice Studio and a lighter Lite cloning app. Native runtimes cover CUDA, MLX for Apple Silicon, and a C++/ggml build for CPU, CUDA, Metal and Vulkan, with GGUF weights. The card also names llama.cpp, Ollama, LM Studio, Jan, Docker Model Runner and Lemonade as local options.

For servers, SGLang-Omni exposes an OpenAI-compatible /v1/audio/speech endpoint on NVIDIA GPUs with streaming and voice cloning. Zyphra also offers a hosted version, Zyphra Cloud; the card gives no pricing for it.

On its own benchmark, ZTTS2-Eval, which spans 55 language subsets, Zyphra reports a median time to first audio of 80 ms with 8 concurrent requests on one H100 80GB, and peak throughput of 457 seconds of audio per second. Quality is scored by combining speaker similarity and word error rate, but the card defers the full results to a blog post and a technical report (arXiv 2611.16525).

A hybrid design that keeps most layers cheap

The notable engineering choice is the sequence mixer. Of the 11 layers, 9 are Gated DeltaNet, a linear-attention design, and only 2 are global attention layers with 16 heads and 4 key-value heads. Linear layers keep a fixed-size state instead of a growing cache, which matters when a model must hold minutes of audio tokens. Zyphra has not said, in the material read for this article, why it chose this ratio.

The idea is not new to language models. Write-ups of Alibaba’s Qwen3-Next describe a 3:1 mix of Gated DeltaNet and full-attention layers, and Sebastian Raschka’s notes explain that an occasional full-attention layer restores precise recall that a compressed recurrent state loses. ZONOS3-Mini leans harder on the cheap layers, at roughly 4.5 to 1. Applying that recipe to a speech model with a 15 Hz codec is what the card presents as new, though the card itself offers no comparison against an all-attention variant.

The sparse routing works in the same direction: with only about 100M parameters active per token, the model reads far less weight from memory at each step than its 800M total suggests. That is consistent with the latency figure Zyphra reports, but the figure comes from the company’s own test setup and has not been independently reproduced.

The runtime list also connects to the wider local-inference ecosystem. The llama.cpp team recently released version 1.0 of llama.cpp and ggml, which it says supports more than 150 architecture variants, and ZONOS3-Mini is among the models arriving through that stack as GGUF weights.

What the card does not settle

For readers who want to try it, the weights are free to download and use commercially under Apache 2.0, subject to the usual license terms. The card shows about 13,068 downloads last month and roughly 301 likes, a sign of early community interest rather than a measure of quality.

Several things remain open. The page gives no numeric quality scores, no comparison table with other open or commercial speech systems, and no explicit release date. It also does not describe training data, or what safeguards exist against cloning someone’s voice without consent, a question that applies to any model that can copy a voice from a one-minute clip. Those details sit in the technical report and blog post, which this article did not review. Until independent tests appear, the speed and quality claims are Zyphra’s own.