Supertonic 3
A 99M on-device speech model for 31 languages that runs through ONNX Runtime. OpenRAIL-M weights; development ended in 2026.
In short. Supertonic 3 reads text aloud in 31 languages and runs entirely on your own device, with no graphics card. It comes with ten preset voices, and its license allows commercial use but bans certain misuses. Its maker stopped developing it in 2026, so it will not get fixes.
Summary
Supertonic 3 is an on-device text-to-speech system from Supertone, a South Korean voice-AI company, released in May 2026. It runs through ONNX Runtime with no cloud call, and the Python SDK (pip install supertonic) downloads the model files on first use. Supertone says the public ONNX files hold about 99 million parameters in total. They are four models of 3.7, 36.4, 256.5 and 101.4 MB, which add up to roughly 400 MB and match 32-bit weights. The output is 44.1 kHz audio. Supertonic 3 widened the 5-language Supertonic 2 to 31 languages: English, Korean, Japanese, Arabic, Bulgarian, Croatian, Czech, Danish, Dutch, Estonian, Finnish, French, German, Greek, Hindi, Hungarian, Indonesian, Italian, Latvian, Lithuanian, Polish, Portuguese, Romanian, Russian, Slovak, Slovenian, Spanish, Swedish, Turkish, Ukrainian and Vietnamese. Chinese is not among them. A na language setting lets the model handle text whose language you do not know.
The open weights include ten fixed preset voice styles. Zero-shot voice cloning was offered only through Supertone's hosted services, which are not part of the open release. The model card says that, compared with Supertonic 2, version 3 makes fewer repeat and skip errors, matches speakers more closely and accepts simple inline expression tags. It also says Supertonic 3 stays within a competitive error range of much larger open models such as VoxCPM2, while using less memory and running fast on a CPU even when compared with larger models measured on an A100 GPU. The card shows these claims as charts rather than tables, and they are the developer's own.
The project has been wound down. Supertone merged a development-discontinuation notice in July 2026 and archived its GitHub repository on 9 September 2026, saying that no updates, bug fixes, security patches or support will be provided. The code and the weights now sit under a supertone-oss-archive GitHub organization and Hugging Face namespace, which the archive names as the place to download the models. The original Supertone/supertonic-3 repository on Hugging Face still exists and has the larger download count, about 93,000 over the 30 days to 11 November 2026.
The weights use the BigScience OpenRAIL-M license, which allows commercial use but requires every copy and derivative to carry its list of banned uses. The sample code is MIT. Supertonic 3 suits a product that needs free, private, multilingual speech from a small download and can live without further updates. Teams that want a maintained engine or voice cloning should compare Pocket TTS and Phonon.
Good for
- Private, offline speech on laptops, phones, browsers and small boards, with no GPU
- Products that need many languages in one small model, including Korean, Japanese, Arabic and most European languages
- Commercial use, as long as the OpenRAIL-M use restrictions are passed on to your users
- Fast synthesis with low memory use on an ordinary CPU
Watch out for
- Development and support have ended and the repository is archived, so expect no fixes or new languages
- The license is OpenRAIL-M with a banned-use list rather than Apache 2.0 or MIT, and every copy must carry the restrictions
- Only ten preset voices are included, and cloning needed Supertone's hosted services
- There is no Chinese
- The quality and speed claims come from the developer's charts and we have not reproduced them
- The original Hugging Face repository may be removed, so use the archive namespace for anything long-lived
Performance claims come from the developer and have not been independently verified.
How much memory does it need?
Memory figures for each variant of Supertonic 3. Image, video, speech and music pipelines also load a text encoder and a decoder, so the memory needed can be larger than the main checkpoint.
| Variants | Parameters | Download size | Memory needed | Fits on |
|---|---|---|---|---|
| Supertonic 3 (ONNX) FP32 | 99M | 0.4 GB | 0.46 GB | Laptop or phone |
Download size covers every component needed to run the model, such as the generator, text encoder and decoder. Memory needed is a stated or published measured minimum where one exists; otherwise it is the download size plus about 15%. Larger images, longer clips and higher resolutions need more.
The license in plain words
The standard Open RAIL-M license written by BigScience in August 2022. It grants free, worldwide copyright and patent rights, including hosting the model as a service, but ties every copy and every derivative to a list of banned uses such as breaking the law, impersonating people without consent, giving medical advice and fully automated decisions about people's legal rights.
- Attribution or model name must be shown
- Acceptable-use policy applies
See also
- Open-weight AI models made in South Korea
- Best open-source text-to-speech models in 2026
- Open speech models: transcription is open, text-to-speech is mixed
- Tiny text-to-speech models that run on a CPU: four open options compared
Frequently asked questions
- What is Supertonic 3?
- Supertonic 3 reads text aloud in 31 languages and runs entirely on your own device, with no graphics card. It comes with ten preset voices, and its license allows commercial use but bans certain misuses. Its maker stopped developing it in 2026, so it will not get fixes.
- Can I use Supertonic 3 commercially?
- Yes, with conditions. Supertonic 3 is released under the OpenRAIL-M, which allows commercial use but adds extra rules: Attribution or model name must be shown; Acceptable-use policy applies. Read the license text before you ship a product.
- Is Supertonic 3 open source?
- Supertonic 3 is open-weight, not open source in the strict sense. You can download the model files, but the OpenRAIL-M adds conditions that the Open Source Initiative definition does not allow.
- Can I run Supertonic 3 on my own computer?
- Yes. The smallest variant, Supertonic 3 (ONNX), needs about 0.46 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
- When was Supertonic 3 released?
- Supertonic 3 was released by Supertone on May 2026.
Official links
Sources
Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.
- Supertonic 3 model card (Hugging Face) huggingface.co
- Supertonic 3 archived model repository (Hugging Face) huggingface.co
- Supertonic repository, archived on 9 Sep 2026 (GitHub) github.com
- Supertonic 3 license text, BigScience Open RAIL-M (Hugging Face) huggingface.co
Release date and facts last verified November 11, 2026.
Related models
| Model | Released | Parameters | Context | License | Smallest variant needs |
|---|---|---|---|---|---|
| Phonon 31M Gradium | 31M | – | Apache 2.0 (Permissive) | 0.0747 GB Laptop or phone | |
| Phonon 40M Gradium | 40M | – | CC BY-NC (Non-commercial) | 0.0858 GB Laptop or phone | |
| Pocket TTS Kyutai | 18M – 100M | – | CC BY 4.0 (Permissive) | 0.127 GB Laptop or phone | |
| Kokoro hexgrad | 82M | – | Apache 2.0 (Permissive) | 0.38 GB Laptop or phone |