# Zyphra/ZONOS3-Mini · Hugging Face — retained record

> Retained reddit record from reddit:LocalLLaMA.

Canonical: https://highsignal.app/data/records/6b470b36126bf749

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:LocalLLaMA)  published 2026-11-11 07:09 ingested 2026-11-12 02:42

[original source ↗](https://www.reddit.com/r/LocalLLaMA/comments/1xtb9gw/zyphrazonos3mini_hugging_face/)

## retained text



```
ZONOS3-Mini is Zyphra’s compact, open-source text-to-speech model, delivering natural reliable speech and high-fidelity voice cloning with just 100 million active parameters. It runs in real time on phones and laptops, including directly in the browser.

Clone a voice from an audio sample, or design a character by shaping age, accent and vocal texture. Steer emotion, pace and delivery down to individual words for games, narration and voice assistants.

Run ZONOS3-Mini locally with our browser and native runtimes, or access it through [Zyphra Cloud](https://cloud.zyphra.com/models/zyphra/ZONOS3-Mini). Model weights are released under Apache 2.0.

**Model details, samples and evaluations are in the** [**blog post**](https://www.zyphra.com/our-work/zonos3-mini)**.** **A hosted version is available at** [**Zyphra Cloud**](https://cloud.zyphra.com/models/zyphra/ZONOS3-Mini)**.**

# [](https://huggingface.co/Zyphra/ZONOS3-Mini#languages)Languages

ZONOS3-Mini supports 43 languages.

|Region|Languages|
|:-|:-|
|Western and Southern Europe|English, French, Spanish, German, Italian, Portuguese, Dutch, Greek|
|Northern Europe|Swedish, Danish, Norwegian, Finnish|
|Central and Eastern Europe|Polish, Czech, Slovak, Hungarian, Romanian, Bulgarian, Russian, Ukrainian|
|Middle East and Caucasus|Turkish, Arabic, Hebrew, Georgian|
|East Asia|Mandarin Chinese, Japanese, Korean|
|Southeast Asia|Vietnamese, Indonesian, Malay, Thai, Tagalog|
|South Asia|Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Odia, Urdu|

# [](https://huggingface.co/Zyphra/ZONOS3-Mini#quick-start)

# Native runtimes

The [ZONOS3-Mini GitHub repository](https://github.com/Zyphra/ZONOS3-Mini) holds Zyphra's own runtimes, which share one HTTP API and serve the Zonos Voice Studio: a CUDA server for NVIDIA GPUs, an MLX server for Apple Silicon, a C++/ggml server and CLI (CPU, CUDA, Metal or Vulkan), and the in-browser WebGPU build. They load the GGUF files in [`gguf/`](https://huggingface.co/Zyphra/ZONOS3-Mini/tree/main/gguf) and the WebGPU shards in [`web/`](https://huggingface.co/Zyphra/ZONOS3-Mini/tree/main/web); the 27 default voices, grouped by language, are in [`gguf/voices/`](https://huggingface.co/Zyphra/ZONOS3-Mini/blob/main/gguf/voices/README.md).

# [](https://huggingface.co/Zyphra/ZONOS3-Mini#architecture-rough-outline)Architecture (rough outline)

Text is NeMo-normalized (English, Mandarin, Japanese, Korean, German, Spanish, French, Italian, Portuguese; other languages pass through as written) and encoded with an 8,192-entry SuperBPE tokenizer. The language model autoregressively generates 15 Hz dual-codec frames (10 codebooks of 4,096 entries each: one semantic codebook distilled from the Wav2Vec2-BERT teacher plus nine acoustic codebooks), conditioned on the text, a speaker prompt and three bucketed acoustic controls. A causal multi-resolution iSTFT head renders the frames to 44.1 kHz audio as they are produced.



||ZONOS3-Mini|
|:-|:-|
|Backbone|11 layers, model dim 1,024; 800M parameters total, about 100M active per token (latent MoE: 64 experts, top-2 routing, from layer 1)|
|Sequence mixer|pattern `gdn3_global1`: 9 Gated DeltaNet (GDN) layers + 2 global attention layers (16 heads x 32 dims, 4 KV heads), attention residuals (`full_attnres`); context 32,768 positions|
|Text|SuperBPE, 8,192 entries (+4 specials), NeMo text normalization at inference|
|Audio tokens|15 Hz, 44.1 kHz dual-encoding codec, 10 x 4,096 codebooks (1 semantic + 9 acoustic); hop 2,940 samples; up to 10 minutes of audio per generation|
|Speaker prompt|raw reference clip (encoded by the codec into a token prefix, up to 60 s, with speaker CFG) **or** a single 192-d MECT-B2 speaker embedding (`ownmect_s2`)|
|Controls|loudness (LUFS, 10 buckets), bandwidth f\_max (Hz, 10 buckets), **speaking rate (UTF-8 bytes per second, 30 buckets)**; each row has an "unknown" index; bucket edges in `params.json` (`condition_edges`)|
|Decoder|causal multi-resolution iSTFT head: `full_mr` (default) or `keep2_plain` (pruned, about 3x faster decode)|
|Precision|bf16 weights (`model.pth`); decoder heads run in fp32|

* **GitHub** : [https://github.com/zyphra/zonos3-mini](https://github.com/zyphra/zonos3-mini)
* **Source** : [https://www.zyphra.com/our-work/zonos3-mini](https://www.zyphra.com/our-work/zonos3-mini)
* **Project** : [https://www.zyphra.com/models/zonos3-mini](https://www.zyphra.com/models/zonos3-mini)
```



document fetched 2026-11-12 00:44 · canonical https://reddit.com/r/LocalLLaMA/comments/1xtb9gw/zyphrazonos3mini_hugging_face
