Created by the independent researcher hexgrad and released with an accompanying inference library.

Kokoro-82M is a lightweight 82 million parameter open weight text to speech model praised for quality relative to its tiny size.

Weights are distributed on the Hugging Face Hub at hexgrad/Kokoro-82M with an inference library on GitHub at hexgrad/kokoro.

Released in late 2024 and quickly became one of the most downloaded open TTS models.

It offers near state of the art speech quality that runs cheaply on CPU or a small GPU, unlike larger proprietary TTS systems.

Built on a StyleTTS2 and ISTFTNet based architecture and runs via the kokoro Python package or ONNX.

Apache 2.0 licensed, open weights, community maintained, roughly 8.3k GitHub stars on the inference repo and millions of monthly Hugging Face downloads.