Kokoro

An 82M-parameter Apache 2.0 speech model with 54 preset voices in 8 languages. Hugely popular and fast, but it cannot clone voices.

In short. Kokoro is a very small text-to-speech model that reads text aloud in 54 ready-made voices across eight languages. It is light enough for a laptop processor, the Apache 2.0 license lets you use it in products, and it is the most downloaded open voice model we track. It cannot copy a new voice from a sample, and its last model release came in February 2025.

Summary

Kokoro is an open text-to-speech model trained and released by hexgrad, an independent developer. The first version appeared on 25 December 2024. Version 1.0, released on 27 January 2025, has 82 million parameters and ships as a 327 MB PyTorch checkpoint plus one small file per voice, a size that matches 32-bit weights. It builds on the StyleTTS 2 design with an ISTFTNet vocoder and, in the author's words, is decoder-only: there is no diffusion step and no voice encoder was released. Text is turned into phonemes by the author's misaki library, with espeak-ng as a fallback, and the kokoro Python package wraps both. Hugging Face counted about 12.5 million downloads over the 30 days to 11 November 2026, far more than any other speech model on this site.

Version 1.0 has 54 voices in eight languages: American and British English, Japanese, Mandarin Chinese, Spanish, French, Hindi, Italian and Brazilian Portuguese, with the card counting the two English accents as one language. The voice page grades every voice for the quality and amount of its training data. Most non-English voices are thin, French has a single voice, and the author warns that support outside English may be weak because of the phonemizer and a lack of training data. A separate v1.1-zh checkpoint, published on 26 February 2025, adds about 100 Chinese voices for 103 voices in two languages. The author says it is not a strict upgrade, because it drops many of the older voices.

The training data is unusual for how openly it is described. Kokoro was trained on a few hundred hours of audio that the developer calls permissive or non-copyrighted: public-domain audio, audio under licenses such as Apache and MIT, and synthetic speech generated by closed text-to-speech models from large providers, with none from open models or voice clones. Two datasets that require attribution, Koniwa and SIWIS, are credited on the card. The author puts the whole training cost at about US$1,000 for 1,000 A100 80GB hours. These are the developer's own figures. Kyutai's January 2026 laptop tests found Kokoro running faster than real time on a CPU, one of only two models in its comparison that did.

The trade-off is flexibility. The voices are fixed, because no encoder was released and the model cannot clone a voice from a sample. The voice page says most voices work best on text of roughly 100 to 200 tokens, are weak on very short text and can sound rushed beyond about 400 tokens, so long text should be split into chunks. There has been no new model release since February 2025, and the author is an individual rather than a company.

Good for

  • Fast narration, read-aloud features and assistant voices where a fixed set of voices is enough
  • Running on a CPU, in a browser or on a small server, because the model has only 82 million parameters
  • Commercial products, since the weights are Apache 2.0 and the card describes the training data
  • English text-to-speech, where the best-graded voices are American English

Watch out for

  • It cannot clone a voice, and you are limited to the preset voices
  • Quality outside English varies a lot, and French has only one voice, so listen to your language before committing
  • Very short or very long inputs can sound weak or rushed, so split long text into chunks
  • Part of the training audio was synthetic speech from closed text-to-speech services, which the developer treats as free of copyright, so check that statement if provenance matters to your legal team
  • It needs a phonemizer, and espeak-ng may need to be installed separately
  • The developer is an independent author with no company behind the model, so do not count on support

Performance claims come from the developer and have not been independently verified.

How much memory does it need?

Memory figures for each variant of Kokoro. 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
Kokoro-82M v1.0 FP32 82M 0.33 GB 0.38 GB Laptop or phone
Kokoro-82M v1.1-zh (Chinese and English) FP32 82M 0.33 GB 0.38 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

One of the most widely used permissive licenses. You may use, change and sell products built on the model with only light conditions.

More about Apache 2.0

See also

Frequently asked questions

What is Kokoro?
Kokoro is a very small text-to-speech model that reads text aloud in 54 ready-made voices across eight languages. It is light enough for a laptop processor, the Apache 2.0 license lets you use it in products, and it is the most downloaded open voice model we track. It cannot copy a new voice from a sample, and its last model release came in February 2025.
Can I use Kokoro commercially?
Yes. Kokoro is released under Apache 2.0, a permissive license. You can use it commercially if you follow the standard conditions, such as keeping the license notice.
Is Kokoro open source?
Kokoro is released under Apache 2.0, a license approved by the Open Source Initiative. Training data and training code may not be public, so it is best described as open weights under an open-source license.
Can I run Kokoro on my own computer?
Yes. The smallest variant, Kokoro-82M v1.0, needs about 0.38 GB of memory at 4-bit precision, which fits a modern laptop or a mid-range graphics card.
When was Kokoro released?
Kokoro was released by hexgrad on January 27, 2025.

Official links

Sources

Specifications come from these primary sources. Benchmark claims, where mentioned, are reported by the developer.

Release date and facts last verified November 11, 2026.

Related models

Model Released Parameters Context License Smallest variant needs
Pocket TTS Kyutai 18M – 100M – CC BY 4.0 (Permissive) 0.127 GB Laptop or phone
Phonon 31M Gradium 31M – Apache 2.0 (Permissive) 0.0747 GB Laptop or phone
Supertonic 3 Supertone 99M – OpenRAIL-M (Conditional) 0.46 GB Laptop or phone
MOSS-TTS-Nano OpenMOSS 100M – Apache 2.0 (Permissive) 0.276 GB Laptop or phone