Model reference · open weights

kokoro-coreml

Not for GPU servers Audio kensora Text→speech 1 build Open weights 1k dl/mo

kokoro-coreml is an open-weight audio or speech model from kensora. Built for Apple devices (Core ML). It does not run on NVIDIA GPUs.

All audio & music

kokoro-coreml is a text-to-speech model developed by kensora for on-device neural synthesis in the Kensora iOS app.
It re-hosts public Kokoro-82M weights as staged Core ML models, including components for the synthesis pipeline and a grapheme-to-phoneme speller.
The model is distributed under the apache-2.0 licence.

Summary of the kensora/kokoro-coreml model card, 2026-10-01

What it is

What it runs on

Memory and cards for kokoro-coreml (BF16)

329 MBweights, file size

1.6 GBruntime overhead

From the model card

Read the model card

On-device neural TTS for the Kensora iOS app (K509, ADR 0085). Public Kokoro-82M weights re-hosted as staged Core ML models, MLX-free.

KokoroFrontA.mlpackage, KokoroProsodyB.mlpackage, KokoroDecoder.mlpackage — the staged synthesis pipeline (front/duration, prosody/F0N, iSTFTNet vocoder).
BartEncoder.mlpackage, BartDecoder.mlpackage — misaki's out-of-vocabulary grapheme-to-phoneme speller (tiny BART), for pronouncing words outside the lexicon.

Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.

How it works

Running it

Built for Apple devices (Core ML). It does not run on NVIDIA GPUs.

Details