ZONOS3-Mini 43-Language Text-to-Speech Model: File List

ZONOS3-Mini 43-Language Text-to-Speech Model: File List

Sample outputs are available on the model card.

At a Glance

Item Value
Repository Zyphra/ZONOS3-Mini
Family guide ZONOS3 guide (1 articles)
Published 2026-11-09
License apache-2.0
Formats safetensors / GGUF / PyTorch (pth) / PyTorch (bin)
Source type Primary source (the publisher itself)

Values determined by this site’s code when the information was collected. Dates are JST.

Overview

ZONOS3-Mini is an open-source text-to-speech (TTS) model released by Zyphra. It outputs audio given text input and also supports voice cloning from reference audio. According to the model card, it operates with 100 million active parameters and is reported to run in real time on smartphones, laptops, and browsers.

It is reported that you can design voices by specifying age, accent, and voice quality, as well as adjust emotions, speaking speed, and intonation on a per-word basis. Intended use cases include games, narration, and voice assistants. This information is based on the official model card.

Specifications

  • Parameter count: Total of 800 million, with approximately 100 million (100M) active per token. It adopts latent MoE (64 experts, top-2 routing, applied from layer 1).
  • Architecture: 11 layers with a model dimension of 1,024. It is a hybrid of 9 layers of Gated DeltaNet (GDN) and 2 layers of global attention, with a context of 32,768 positions.
  • Text processing: Uses the SuperBPE tokenizer with 8,192 entries. Text normalization is performed via NeMo during inference. Normalization targets English, Chinese, Japanese, Korean, German, Spanish, French, Italian, and Portuguese, while other languages are passed through as-is.
  • Audio tokens: 15Hz, dual-encoding codec (ZAC 1) at 44.1kHz. Composed of 10 codebooks with 4,096 entries each (1 semantic, 9 acoustic). Can output up to 10 minutes of audio in a single generation.
  • Decoder: Causal multi-resolution iSTFT head. The default is full_mr, and the pruned variant keep2_plain is reported to provide about 3x faster decoding.
  • Output: 44.1kHz audio. It also supports streaming output.
  • Speaker prompts: Uses either a reference clip (up to 60 seconds, with speaker CFG) or a single 192-dimensional speaker embedding.
  • Control parameters: Volume (LUFS, 10 levels), bandwidth f_max (10 levels), and speaking speed (UTF-8 bytes per second, 30 levels).
  • Supported languages: 43 languages. Includes European languages as well as Arabic, Hindi, Tamil, Vietnamese, Thai, Japanese, Korean, and Mandarin Chinese.
  • Precision: Weights operate in bf16, and the decoder head operates in fp32.
  • Licensing terms: Model weights are released under Apache 2.0.
  • Default voices: 27 built-in voices are included.

Performance and Quality

姉妹モデルのモデルカードには、H100 80GB 1枚での品質と速度の比較図が掲載されています。However, numerical tables are not present in the card body, and the names of other compared models or scores could not be verified from the documentation. The following are values published by Zyphra themselves.

  • Quality metric: A combined value of speaker similarity (macro) and WER (macro) on a 55-language subset of ZTTS2-Eval. It targets FLEURS-R read speech (35 languages) and VoxBlink2 in-the-wild recordings (20 languages), and ZONOS3-Mini is measured using speaker embedding prompts.
  • Time to first audio (median): 80 milliseconds across 8 concurrent requests.
  • Peak throughput: 457 seconds of audio per second.

WER stands for Word Error Rate, where lower is better. This quality metric combines speaker similarity and WER. Detailed results including the ZTTS2 capability index are reported to be published on the official blog. Superiority or inferiority compared to other models cannot be determined from the materials in the text alone.

Strengths and Use Cases

Use cases listed in the model card include games, narration, and voice assistants. In addition to cloning voices from reference audio, it is reported that character voices can be designed by specifying age, accent, and voice quality. Emotions, speaking speed, and intonation can be adjusted on a per-word basis.

Tags include text-to-speech, voice-cloning, and multilingual, supporting 43 languages. It can output up to 10 minutes of audio in a single generation and supports streaming output, making it suitable for longer narrations and conversational use cases where low latency is required.

Furthermore, it is reported to run in real time on smartphones, laptops, and browsers, making it well-suited for use cases that need to run locally without setting up a server.

Distributed Files

Weight files published in Zyphra/ZONOS3-Mini, listed by this site from the Hugging Face API. Sizes are the actual file sizes.

File Size
autoencoder/generator.pt 7.96GB
autoencoder/heads/full_mr/step_0050000.pt 4.01GB
autoencoder/heads/keep2_plain/step_0050000.pt 518MB
autoencoder/teacher/model.safetensors 774MB
gguf/edit/accents.gguf 257MB
gguf/spk-encoder.gguf 45MB
gguf/ssl-teacher.gguf 1.55GB
gguf/steer.gguf 0MB
gguf/zac-encoder.gguf 4.03GB
gguf/zac-keep2.gguf 452MB
gguf/zac.gguf 2.20GB
gguf/zonos3-ptq-edit.gguf 663MB
gguf/zonos3-ptq.gguf 663MB
lm/zonos3_mini/model.pth 1.77GB
speaker_encoders/ownmect_s2/ckpt_0037.pth 50MB
web/aligner/shard-000.bin 58MB
web/aligner/shard-001.bin 61MB
web/aligner/shard-002.bin 66MB
web/aligner/shard-003.bin 61MB
web/aligner/shard-004.bin 65MB

88 more file(s) are listed on Hugging Face.

How to Get It

The model is available on Hugging Face at Zyphra/ZONOS3-Mini. There is no gating (agreement to terms of service required), so it can be acquired without consent. Distribution formats include safetensors, GGUF, and PyTorch (pth / bin). GGUF files are located under gguv/, and shards for WebGPU are under web/. The 27 default voices are included in gguv/voices/.

Trying It in the Browser

If you use a WebGPU-compatible browser, you can run it on your own GPU with no installation or server required. The model is downloaded upon initial access and cached by the browser thereafter.

  • Zonos Voice Studio: A full-featured studio.
  • Lite: Choose, upload, or record a voice to immediately start speaking with a cloned voice.

Running with SGLang-Omni

For NVIDIA GPUs, it can be served via an OpenAI-compatible /v1/audio/speech endpoint. It supports streaming and voice cloning from reference clips (no transcription required). Follow the SGLang-Omni installation guide for setup. Start the server with the following command:

sgl-omni serve --model-path Zyphra/ZONOS3-Mini --port 8000

Once curl http://localhost:8000/health returns "healthy", you can generate audio using the default voices.

curl -X POST http://localhost:8000/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"input": "Hello from ZONOS3-Mini!"}' \
  --output output.wav

To clone a voice, add "references": [{"audio_path": ""}]. To stream 44.1kHz audio, add "stream": true, "response_format": "pcm". Details on cloning and sampling options can be found in the ZONOS3 cookbook.

Native Runtimes

The ZONOS3-Mini GitHub repository provides Zyphra’s own runtime. It features a common HTTP API and is also used in Zonos Voice Studio.

  • CUDA server for NVIDIA GPUs
  • MLX server for Apple Silicon
  • C++/ggml server and CLI (CPU, CUDA, Metal, Vulkan)
  • In-browser WebGPU build

Zyphra Cloud is also available as a hosted option.

Can You Run It Locally?

The publisher distributes this model as safetensors / GGUF / PyTorch (pth) / PyTorch (bin).

License — apache-2.0 (Commercial use allowed): Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.

Compiled by this site’s code from the published formats and the license field. License summaries are not legal advice — check the publisher’s original terms before relying on them.

Quantized and Converted Variants

→ Scroll horizontally to see all columns

Added Publisher Format Repository Smallest VRAM tier (build, est. memory)
2026-11-16 ggml-org GGUF ggml-org/ZONOS3-Mini-GGUF BF16 1.8GB (fits in 4GB VRAM)

File sizes of each build:

  • Available builds in ggml-org/ZONOS3-Mini-GGUF: Q8_0 0.8GB / BF16 1.5GB

In addition, 5 converted build(s) from other uploaders exist on Hugging Face; this site lists only builds from the model’s publisher or established quantization maintainers.

This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files. What the quantization names mean: glossary.

Other Models for the Same Task

Recent audio models covered by Local Model Watch, newest first. Grouped by the task each publisher declares on Hugging Face (pipeline_tag); the smallest VRAM tier is this site’s estimate.

See all audio models →

What to Read Next

Sources