Skip to main content
株式会社オブライト
株式会社オブライト

ZONOS3-Mini: An Open 43-Language TTS with 100M Active Parameters

Requirements, Japanese Support and Voice Cloning [vs Irodori-TTS and Fish Audio]

Published 2026/11/1712 min read

ZONOS3-Mini is an open-weight TTS model from Zyphra with about 800M total and 100M active parameters. It supports 43 languages including Japanese and voice cloning, with Apache 2.0 weights. We cover official requirements and real file sizes for browser, NVIDIA, Mac and CPU, how to run it, and how it differs from Irodori-TTS, Fish Audio, VoiceStudio and ElevenLabs. Updated Nov 2026.

ZONOS3-Mini is an open-weight text-to-speech (TTS) model from Zyphra, with about 800M total and about 100M active parameters per token (MoE), support for 43 languages including Japanese, and voice cloning from an audio sample. The weights are released under Apache 2.0 and are free to use. The official model card says it runs in the browser (WebGPU), on NVIDIA GPUs, on Apple Silicon, and on CPU (C++/ggml), and memory needs differ a lot between these. This article puts the officially stated requirements and the real file sizes first.

This article is based on the Hugging Face model card, Zyphra's official blog post (published November 9, 2026), the GitHub repository, and the arXiv technical report (2611.16525, submitted November 12, 2026), as of November 16, 2026. The Hugging Face repository was created on November 9, 2026, and as of November 16, 2026 it showed 39,862 downloads and 393 likes, and the GitHub repository had 166 stars. The weights are under Apache 2.0 and the GitHub code is under MIT. All evaluation results are figures published by Zyphra, not independently verified.

Why it is getting attention now

The reason is its light weight, with only 100M active parameters, while still offering multilingual speech and voice cloning. According to the official blog, in the browser (WebGPU) the time to first audio is under 570 ms on a Pixel 9 and under 150 ms on an M3 MacBook Air. These are Zyphra's published figures, and we could not confirm the measurement conditions. The model card says that on one H100 80GB, the median time to first audio with 8 concurrent requests is 80 ms and peak throughput is 457 seconds of audio per second.

On evaluation, the official blog says ZONOS3-Mini ranks first among open models on the Capability Index of Zyphra's own benchmark ZTTS2-Eval (35 languages of FLEURS-R read speech plus 20 languages of VoxBlink2 in-the-wild recordings, 55 subsets in total), followed by Fish S2 Pro and MOSS TTS v1.5, but no score is given in the text. The arXiv abstract reports a macro-average WER of 6.71% over the 55 subsets and calls it competitive with larger, slower state-of-the-art models. Training took 2.5 days on 112 H100 GPUs, and the audio codec was replaced with ZAC-1 (15 Hz). All of these are Zyphra's published figures.

As for earlier generations, we found no mention of Zonos v0.1 on the official pages. The official blog compares ZONOS3-Mini with ZONOS2 (8B parameters, DAC codec at 86 Hz) and says the new model has 800M total parameters, a custom codec, a hybrid Gated DeltaNet and attention backbone, and a new speaker encoder.

What it can do

FeatureDetails
Multilingual TTS43 languages, including Japanese, Mandarin Chinese, Korean, English and Hindi. Output is 44.1 kHz mono
Voice cloningReproduces a speaker from a reference clip (up to 60 seconds). Two modes: a "raw reference" encoded by the codec, or a 192-dimensional speaker embedding. SGLang-Omni needs no transcript
Voice design and controlShape age, accent and vocal texture to design a voice. The model card says emotion, pace and delivery can be steered down to individual words. It also has three control values: loudness, bandwidth and speaking rate
StreamingReturns audio while generating. Provides an OpenAI-compatible /v1/audio/speech and its own /v2/generate
LengthUp to 10 minutes of audio per generation (architecture table in the model card)
Speech editingPartial editing using a transcript. However, the README says the required LoRA is in a private repository, so whether it is usable as-is is unconfirmed

Requirements and how to run it

First, the table below summarizes requirements and real file sizes per environment. Sizes are byte counts from the Hugging Face API (retrieved November 16, 2026) converted to decimal GB, so they differ slightly from the rounded figures in the official README. The official sources do not state a minimum VRAM or RAM for any environment. Note that the file sizes in the table are not lower bounds for the memory needed at run time.

How to runOfficially stated requirementsReal file sizesNot stated officially
Browser (WebGPU)A WebGPU-capable browser. No install; the weights download on the first visit and the browser caches themThe whole web/ folder is about 5.69 GB (the README says about 5.3 GB, of which the full-size decoder zac-full_mr is about 1.23 GB). The first load fetches backbone and zac, about 1.0 GB (README). Using an uploaded voice adds encoder and teacher, about 2.8 GB (README)Required memory. The blog reports under 570 ms to first audio on a Pixel 9
NVIDIA CUDA (SGLang-Omni / Zyphra CUDA server)Linux x86_64 with an NVIDIA GPU (tested on H100), CUDA 12.8 toolkit, Python 3.12 (managed by uv). The first start JIT-compiles kernels and takes several minutes. On pre-Hopper GPUs such as the A100, portable Triton kernels are selectedThe README says the PyTorch weights are about 9 GB (default decoder). The LM model.pth is 1.77 GB (bf16). The whole repository is about 30.7 GBMinimum VRAM. How widely it was tested beyond H100
Apple Silicon (MLX server)An Apple Silicon Mac. uv sync and a GGUF download. Defaults to 127.0.0.1:1919The README says the gguf/ folder is 9.3 GB (about 9.9 GB measured on Hugging Face). The minimum set is backbone-ptq about 0.66 GB + zac-keep2 about 0.45 GB + spk-encoder about 0.05 GB, about 1.16 GB in totalRequired RAM. The blog reports under 150 ms to first audio on an M3 MacBook Air (the runtime used is not stated)
CPU and others (C++/ggml server and CLI)Supports CPU, CUDA, Metal and Vulkan. Needs CMake and a Rust toolchain (used to build the text-normalization library). The README says Metal, Vulkan and Windows builds depend on the installed toolchain and backend operatorsThe same GGUF files as MLX. The default backbone is post-training quantized (Q4_0 experts, Q8_0 mixers). The full-size decoder zac.gguf is about 2.20 GB, zac-encoder.gguf for raw-reference cloning is about 4.03 GB, and ssl-teacher.gguf is about 1.55 GBRequired RAM. CPU speed (real-time factor) could not be confirmed

As the table shows, the smallest ways to try it are the browser build and the minimum GGUF set of about 1.16 GB. By contrast, for NVIDIA GPU servers the only statement is that it was tested on H100, and whether it runs on a 24 GB gaming GPU is not stated officially. Also, the preset voices are grouped by language folder, with English, French, German, Korean, Chinese and others, but there is no Japanese folder. To make a Japanese voice you need to prepare your own reference audio.

The weights on Hugging Face are under Apache 2.0 (commercial use allowed), and the runtime code on GitHub is under MIT. The GitHub README also notes that the optional phonemizer espeak-ng is GPL-3.0 (opt-in) and that FuseAlign is Apache-2.0. A hosted version is also offered on Zyphra Cloud, but we did not check its pricing.

Quick start

- Try it with no install: open Zonos Voice Studio (https://zyphra.com/zonos3-mini/localplayground) in a browser, or Lite (/lite), which focuses on voice cloning. You need a WebGPU-capable browser and must wait for the first model download.
- Run it as an NVIDIA GPU server: install SGLang-Omni, run sgl-omni serve, and POST to the OpenAI-compatible /v1/audio/speech. For voice cloning, pass a reference clip in references.
- Run it on a Mac: clone the repository, run uv sync --locked --package zonos3-mlx, download the GGUF files, and run servers/mlx/scripts/serve.sh.
- Run it with C++/ggml: build with CMake in servers/ggml, download the three minimum GGUF files, and generate with zonos3-cli.

# 1. NVIDIA GPU(SGLang-Omni をインストール済みの場合。モデルカードの手順)
sgl-omni serve --model-path Zyphra/ZONOS3-Mini --port 8000

# 別のターミナルで、既定の声で生成
curl -X POST http://localhost:8000/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"input": "Hello from ZONOS3-Mini!"}' \
  --output output.wav

# 2. Apple Silicon(GitHub README の手順。クローン先は README では ZONOS3 表記)
git clone --recurse-submodules git@github.com:Zyphra/ZONOS3.git
cd ZONOS3
uv sync --locked --package zonos3-mlx
hf download Zyphra/ZONOS3-Mini --include 'gguf/*' --local-dir servers/ggml
servers/mlx/scripts/serve.sh      # 127.0.0.1:1919

# 3. C++/ggml(servers/ggml で。最小構成の GGUF 3点)
cmake -B build -DCMAKE_BUILD_TYPE=Release && cmake --build build -j
hf download Zyphra/ZONOS3-Mini gguf/backbone-ptq.gguf gguf/zac-keep2.gguf gguf/spk-encoder.gguf --local-dir .
./build/zonos3-cli gguf/backbone-ptq.gguf --tts 'Hello, world.' out.wav --seed 1
./build/zonos3-cli gguf/backbone-ptq.gguf --tts 'Hello again.' clone.wav --clone <参照音声ファイル> --seed 1

For Japanese, the model card lists Japanese among the languages whose input text is normalized with NeMo-based rules (English, Mandarin, Japanese, Korean, German, Spanish, French, Italian and Portuguese; other languages pass through as written). Because there are no official Japanese numbers, whether the reading and prosody actually sound natural is something to check with your own sentences.

How it compares with other TTS models

ItemZONOS3-MiniIrodori-TTS v4-LargeFish Audio Drama 3VoiceStudioElevenLabs
FormOpen weightsOpen weightsOpen weightsLocal app (bundles many engines)Cloud service
ParametersAbout 800M total, about 100M activeAbout 3.29B35B (about 3B active)Depends on the engineNot disclosed
LicenseWeights Apache 2.0 (code MIT)Weights Gemma Terms of Use (code MIT)Fish Audio Community LicenseApp AGPL-3.0 (models have their own)Commercial service
JapaneseAmong its 43 languages. No Japanese-only evaluation is publishedJapanese only83 languages646-language catalog (depends on the engine)Not checked in this article
Voice cloningReference clip up to 60 sReference audio up to 120 s in totalSupportedSupported (default engine works from a few seconds)Starter plan and above (Instant Voice Cloning)
Local runBrowser, NVIDIA, Mac, CPUYes (VRAM not stated officially)Yes (48 GB-class GPU as a guide)Yes (from 8 GB RAM, no GPU needed)No (API and app only)
PriceWeights free (hosted version not checked)FreeFree self-hosting under conditions. API list price $50 per 1M charactersFreeFree tier $0 (10,000 credits/month), Starter $6/month (30,000 credits), Creator $22/month

As a rule of thumb: if Japanese reading and voice quality come first, the Japanese-only Irodori-TTS-v4-Large is a candidate. If you value acting directions and multiple speakers and can prepare a 48 GB-class GPU, there is Fish Audio Drama 3. If you want to switch between several engines in a GUI, VoiceStudio is closer. ZONOS3-Mini is worth considering when you want something light that also runs on phones, laptops and in the browser, or want to cover many languages with one model. ElevenLabs is a cloud option within its price list (here, the free tier plus Starter and Creator as confirmed on the official page), and the deciding factor is whether your data may leave your environment. The figures come from each article or official page and were not measured under the same conditions.

Caveats and who it suits

- Use voice cloning only on voices you have consent for. We found no statement about watermarking or misuse prevention in the official model card, blog or GitHub top page. Treat watermarking of generated audio as not guaranteed by the model, and set consent-recording and disclosure rules on the operations side. Impersonation fraud, such as a payment request that imitates a president's voice, is causing real damage; countermeasures are summarized in countering president-impersonation fraud.
- There are no official numbers for Japanese quality. The official blog and the arXiv abstract show aggregate results over 35 plus 20 languages, but we could not find a Japanese-only value. It is not necessarily at the level of Japanese-only models.
- Parameter counts are stated differently. The model card says 800M total and about 100M active, while the arXiv abstract summary reads as 886M with 100M active. This is probably a difference in what is counted (for example, whether the codec is included), but we found no official explanation.
- Measure the memory yourself. The official sources give no minimum VRAM or RAM. File sizes are only a guide.
- Setup is somewhat involved. The CUDA server spends several minutes JIT-compiling on first start, the README clone URL says ZONOS3 while the repository is named ZONOS3-Mini, and submodules must be fetched.
- Who it suits: developers who want to try a multilingual TTS that runs lightly locally or in the browser, people considering voice output on phones or embedded devices, and anyone looking for a model whose commercial use is easy to check under Apache 2.0. If Japanese quality is the only priority, listen to a Japanese-only model first.

FAQ

Is ZONOS3-Mini free for commercial use?

The weights on Hugging Face are under Apache 2.0, so they are free to use, including commercially. The GitHub code is MIT. The optional phonemizer espeak-ng is GPL-3.0 (opt-in), so check separately if you bundle it. We did not check pricing for the hosted version on Zyphra Cloud.

Does it speak Japanese? How good is it?

Japanese is among the 43 languages. However, we could not find a Japanese-only evaluation in the official blog, the model card or the arXiv abstract. The evaluation is an aggregate over 35 plus 20 languages, with a macro-average WER of 6.71% (Zyphra's published figure). The reliable way to judge Japanese quality is to try your own sentences.

How much VRAM or memory does it need?

No minimum is stated officially. What we know is file size: the minimum GGUF set (backbone-ptq, zac-keep2, spk-encoder) is about 1.16 GB in total, the PyTorch weights are about 9 GB according to the README, and the browser web folder is about 5.69 GB. Run-time memory is higher, so you need to measure it in your own environment.

What audio do I need for voice cloning?

A reference clip of up to 60 seconds. SGLang-Omni needs no transcript. The voice can be used either as a 192-dimensional embedding or encoded by the codec. We found no official recommendation for the ideal clip length or quality (such as the effect of noise or background music).

Does it run on a phone?

The blog says the browser (WebGPU) build reaches first audio in under 570 ms on a Pixel 9. However, the required memory and the range of supported devices are not stated officially.

Summary

ZONOS3-Mini is an Apache 2.0 open TTS that offers 43 languages and voice cloning with about 100M active parameters. The browser build is the entry point for a quick trial, SGLang-Omni for servers, and the minimum GGUF set of about 1.16 GB for Macs and CPUs. On the other hand, required memory, Japanese-only quality and misuse-prevention mechanisms are not described officially, so check them in your own environment and with your own text before adopting it. The evaluation figures are Zyphra's published values, and independent verification remains to come.

Choosing a speech synthesis model and integrating it into internal systems is also something we cover under AI consulting.

References (primary sources)

株式会社オブライト

Oflight Inc. is a Tokyo-based IT company providing software development, digital signage, web development, network systems, and IoT/AR/VR technology services.

Feel free to contact us

Contact Us