ZONOS3-Mini: Run Zyphra’s Open Voice Cloning TTS Locally (43 Languages, VRAM, Setup, API)

Zyphra has released ZONOS3-Mini, an open text-to-speech model that clones a voice from a short clip and speaks 43 languages, including Hindi, Tamil, Bengali and nine more Indian languages.

The surprise is the size. It uses only about 100M active parameters (800M total), yet Zyphra ranks it first among open TTS models. It runs on an NVIDIA GPU, a Mac, a plain CPU, or straight in your browser.

This guide covers the specs, every benchmark number, hardware needs, setup for each runtime, the cloud API, licence, known problems and how it compares with other voice models.

What ZONOS3-Mini Is

Zyphra is a US AI lab known for the original open Zonos TTS models. ZONOS3-Mini is the first model of its third generation, built to be fast and small enough for phones and laptops.

The big change from ZONOS2 is efficiency. ZONOS2 used an 8B backbone and an 86 Hz audio codec. ZONOS3-Mini uses about 100M active parameters and a new 15 Hz codec, so it generates far fewer tokens per second of speech.

You give it text plus a voice. The voice can be a short reference clip (zero-shot cloning), one of 27 built-in voices, or a saved speaker embedding. It then streams 44.1 kHz audio as it goes.

Specs at a Glance

All 43 Languages

This is one of the widest language lists of any open voice cloning model. Indian readers get 12 Indian languages plus Urdu.

Note: Zyphra’s tests cover 35 languages of clean speech and 20 of real-world cloning. Quality on the remaining languages, and on Mandarin tones or Japanese pitch accent, has not been checked independently yet.

Benchmarks and Speed

Zyphra built a new test suite, ZTTS2-Eval. It mixes clean read speech (FLEURS-R, 35 languages) with in-the-wild voice cloning (VoxBlink2, 20 languages), scored by speech recognition for accuracy and by speaker similarity.

Results are combined into one “Capability Index” from 420 test items. All numbers below are vendor-reported.

Speed numbers

Warning: ZTTS2-Eval was designed by Zyphra and has not been reproduced by anyone else yet. Zyphra’s own launch thread also quoted slower browser and laptop latency (under 800 ms and 450 ms) than its blog, so treat the speed numbers as best case.

Hardware: What You Need

Zyphra does not publish official VRAM numbers. The figures below come from its README and early testers, so use them as a guide.

• Browser (WebGPU): any recent Chrome or Edge with WebGPU. About 920 MB downloads on first visit, then it is cached.
• NVIDIA GPU with ggml build: about 2.4 GB VRAM with the quantized backbone (0.66 GB) and fast decoder (0.45 GB). About 4.2 GB with the full 2.2 GB decoder. Any 6-8 GB card is enough.
• NVIDIA CUDA server (Linux): about 15 GB of weights download on first run. Zyphra tested on H100; it is aimed at servers handling many users.
• Apple Silicon (MLX): runs on M-series Macs. One tester saw about 16 GB peak memory during the first voice clone, then normal use afterwards.
• CPU only: works through the ggml build. A tester reported about 1.45 s of compute for a 7.2 s clip.

Tip: on a Windows gaming PC, the ggml build is the simplest route. The CUDA server needs Linux, Python 3.12 and the CUDA 12.8 toolkit.

Try It Free in Your Browser

1. Hugging Face Space. Open huggingface.co/spaces/Zyphra/ZONOS3-Mini. Type text, upload a 5-10 second voice clip or pick a stock voice, and press generate. It runs on shared ZeroGPU, so you may wait in a queue.
2. WebGPU Voice Studio. Open Zyphra’s local playground at zyphra.com/zonos3-mini/localplayground. The model downloads once (about 920 MB) and then runs on your own GPU, so your text and voice clip stay on your device.
3. Check the result. If the voice sounds off, try a cleaner clip: one speaker, no music, no echo, 5-10 seconds long.

Run It Locally, Step by Step

The official code lives at github.com/Zyphra/ZONOS3-Mini (MIT). It uses uv for Python packages and pins Python 3.12. Install uv first from astral.sh.

Linux with an NVIDIA GPU (CUDA server)

1. Install the CUDA 12.8 toolkit. The scripts look for it at /usr/local/cuda-12.8 (set CUDA_HOME if it lives elsewhere).
2. Clone with submodules. git clone --recurse-submodules https://github.com/Zyphra/ZONOS3-Mini.git then cd ZONOS3-Mini.
3. Install packages. uv sync --locked --all-packages
4. Log in to Hugging Face if asked. Run hf auth login or set HF_TOKEN.
5. Start the server. servers/sgl/scripts/serve.sh starts it on 127.0.0.1:1919. The first start takes several minutes while it compiles kernels.
6. Open the studio. Go to http://127.0.0.1:1919/studio for the full Voice Studio, or /lite for a simple voice cloning page.

Apple Silicon Mac (MLX)

1. Clone and install. After cloning as above, run uv sync --locked --package zonos3-mlx
2. Download the GGUF files. hf download Zyphra/ZONOS3-Mini --include 'gguf/*' --local-dir servers/ggml
3. Serve. servers/mlx/scripts/serve.sh and open http://127.0.0.1:1919/

Windows, Linux or CPU (C++/ggml)

The ggml build needs no Python at runtime and supports CPU, CUDA, Metal and Vulkan. Download the same gguf/ folder as above, then follow servers/ggml/README.md in the repo to build the server and command-line tool with CMake.

Other options

• SGLang-Omni (day-0 support): sgl-omni serve --model-path Zyphra/ZONOS3-Mini --port 8000
• Plain PyTorch reference: uv run python -m zonos3.reference --model hf://Zyphra/ZONOS3-Mini/lm/zonos3_mini --text "Hello there" -o ref.wav --device cuda
• Transformers: a separate repo, Zyphra/ZONOS3-Mini-hf, holds a Transformers version.

Using the Local API and Voice Cloning

The local server speaks the same format as OpenAI’s speech API, so many apps that support OpenAI TTS can point at it instead.

curl -X POST http://127.0.0.1:1919/v1/audio/speech \
  -H "Content-Type: application/json" \
  -d '{"model": "zonos3", "input": "Namaste! This voice runs on my own PC.", "response_format": "wav"}' \
  --output hello.wav

For cloning and language control, use the native endpoint POST /v2/generate. Each request needs a voice:

• speaker_audio_base64: your reference clip, base64-encoded.
• speaker_embedding_id: a saved voice; list them with GET /v2/speakers.
• language: for example en_us.
• stream and format: set "format": "wav" with "stream": false for a file. Streaming output is raw float32 PCM at 44.1 kHz mono.

Tip: clone a voice once and reuse its embedding. Testers found cached clones run as fast as the stock voices.

Zyphra Cloud API and Price

If you do not want to host it, Zyphra Cloud serves the same model. At launch, testers reported these terms (check cloud.zyphra.com before you build on them):

Note: pricing is per byte, not per character. Hindi, Chinese, Japanese and other non-Latin scripts use 2-3 bytes per character in UTF-8, so the same sentence costs more than in English.

Voice and Emotion Controls

The Voice Studio exposes “steering” controls on top of plain cloning:

• Speaker: age, gender presentation and texture (soft, raspy, bright, deep).
• Accent: British and Australian English.
• Delivery: pace, emotion (happy, angry, scared), projection, liveliness and whisper.
• Audio: loudness, bandwidth and speaking rate settings.
• Speech editing: change a few words inside an existing recording, using zonos3-ptq-edit.gguf and the Zyphra/fusealign aligner.

Zyphra is honest that these are targets, not guarantees. Pace can also change loudness and pitch, there is no true breathiness control, and age settings are not measured ages.

ZONOS3-Mini vs Other Voice Models

If you mainly want a cheaper paid service instead, see our best ElevenLabs alternatives ranking with prices.

Limits, Licence and Safe Use

• Licence: Apache 2.0 for weights, so you can use it in paid products. The ZAC-1 codec and helper repos carry their own licences; check them before shipping.
• No watermark: Zyphra has not documented audio watermarking or abuse filters.
• Language quality varies: lower-resource languages were scored with Whisper, and per-language scores are not published.
• Parameter count mismatch: the blog says 800M total, the model page shows 886M-0.9B. It does not change how you run it.
• Long-form emotion: not tested independently yet.

Warning: only clone voices you own or have clear permission to use. Copying a real person’s voice without consent can break the law in many countries and the terms of most platforms.

Troubleshooting

• 401 or “repository not found” on download: the README notes some repos were private at launch. Run hf auth login or set HF_TOKEN, and accept any terms on the model page.
• git clone fails for Zyphra/ZONOS3.git: the README uses an older repo name and an SSH URL. Clone https://github.com/Zyphra/ZONOS3-Mini.git instead.
• Missing submodules: run git submodule update --init third_party/fusealign-multilingual servers/ggml/ggml.
• First CUDA start seems frozen: it is compiling kernels and building a text-normalisation cache. Wait several minutes.
• Windows build error about a member named head: a tester hit this with MSVC. Build under WSL2, or use the browser or Space until a fix lands.
• Garbage audio with Japanese text on Windows: Japanese Windows passes CLI text as CP932. Send text through the API as UTF-8 instead of the command line.
• Browser version fails: WebGPU support differs between browsers and drivers. Try the latest Chrome or Edge and update your GPU driver.
• Cloned voice sounds wrong: use a clean 5-10 s clip with one speaker, then tweak the speaker CFG or embedding settings.

My Take

The headline for me is not the benchmark. It is that a model with about 100M active parameters can clone a voice in 43 languages and run in a browser tab. That changes who can build voice apps.

For Indian creators, the list of 12 Indian languages under Apache 2.0 is the standout. Earlier small open models usually stopped at Hindi.

I would still be careful. ZTTS2-Eval is Zyphra’s own test, and nobody has published per-language results yet. Try your own language in the free Space before you commit, and keep a cloud fallback for anything customer-facing.

For local AI basics, our Ollama guide and Whisper guide pair well with it for a fully offline voice assistant.

FAQ

Is ZONOS3-Mini free for commercial use?

Yes. The weights are Apache 2.0 and the runtime code is MIT, so you can use it in paid products. Check the separate licences of the ZAC-1 codec and helper repos.

Does ZONOS3-Mini support Hindi and other Indian languages?

Yes. It supports Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Odia and Urdu. Quality per language has not been published, so test your language first.

How much VRAM does ZONOS3-Mini need?

There is no official figure. Testers saw about 2.4 GB VRAM with the quantized ggml build on an RTX 5080, and about 4.2 GB with the full decoder.

How long should the voice clip be for cloning?

Zyphra’s example uses a 5-10 second clip. The model accepts up to 60 seconds of raw reference audio.

Can I run ZONOS3-Mini without a GPU?

Yes. The C++/ggml build runs on CPU, and the WebGPU version runs in a browser using whatever GPU your device has.

Is it better than ElevenLabs?

Zyphra compared samples with ElevenLabs but published no score against it. It is free and runs locally; ElevenLabs is more polished and needs no setup.

Does it work with apps that use the OpenAI TTS API?

Mostly yes. The local server offers an OpenAI-compatible /v1/audio/speech endpoint with model name zonos3.

Sources

• Zyphra, “ZONOS3-Mini” Hugging Face blog: huggingface.co/blog/Zyphra/zonos3-mini
• Model card: huggingface.co/Zyphra/ZONOS3-Mini
• GitHub README (MIT): github.com/Zyphra/ZONOS3-Mini
• Zyphra post: zyphra.com/our-work/zonos3-mini
• ZTTS2-Eval: zyphra.github.io/ZTTS2-Eval
• Independent review and tests (hardware, cloud pricing): aireiter.com/blog/zonos3-mini-review
• Explainx.ai launch coverage: explainx.ai/blog/zyphra-zonos3-mini-open-tts-apache-webgpu-november-2026