Onyx Kokoro 82M
A mixed-precision MLX quantization of Kokoro-82M (StyleTTS2-derived, Apache-2.0), packaged for Onyx — an on-device iOS AI assistant. Used by OnyxTTS, Onyx's own from-scratch Swift + MLX text-to-speech engine, to...
Model source
- Kind
- Other
- File size
- 510 KB–148 MB
- Base model
- hexgrad/Kokoro-82M
- Version
- vRevision 17a1eb7
- Updated
- Oct 13, 2026
- Downloads
- 106
- Source
- Hugging Face
Source description
Onyx-Kokoro-82M
A mixed-precision MLX quantization of Kokoro-82M (StyleTTS2-derived, Apache-2.0), packaged for Onyx — an on-device iOS AI assistant. Used by OnyxTTS, Onyx's own from-scratch Swift + MLX text-to-speech engine, to speak assistant replies in a natural neural voice fully on-device (Apple's AVSpeechSynthesizer is the fallback).
What this is
- Mixed 4-bit quant (group size 64): the PL-BERT (ALBERT) text encoder's Linear weights are 4-bit; the
entire vocoder path — the prosody predictor (LSTMs + AdaIN convs → pitch/energy) and the iSTFTNet generator —
stays fp16. The vocoder directly shapes the waveform and is quant-sensitive, so it is deliberately NOT
quantized. ~330 MB fp32 → 148 MB (
model.safetensors). - Voice packs (
voices/*.safetensors): 54 voices (af_/am_ American, bf_/bm_ British, plus jf/jm/zf/zm/ ef/em/ff/hf/hm/if/im/pf/pm for other languages). Each is a[N, 1, 256]style tensor indexed by phoneme-sequence length. config.jsoncarries the standard Kokoro config plus aquantizationblock ({group_size: 64, bits: 4}) so the loader dequantizes the right layers.
Provenance
Weights derived from prince-canuma/Kokoro-82M (the mlx-audio mirror of hexgrad/Kokoro-82M) via Onyx's
quantize_kokoro.py. Verified to load + synthesize valid 24 kHz audio end-to-end through OnyxTTS on Apple
Silicon. Apache-2.0, same as the base model.
Sources
1 source
Sources
1 sourceVerified Aug 15
Model artifacts
55 artifacts
Model artifacts
55 artifactsvoices/af_jessica.safetensors
safetensors · 510 KB · SHA-256 93930bbe363c…7003 · Hugging Face
Downloadvoices/am_michael.safetensors
safetensors · 510 KB · SHA-256 19a866143045…a43c · Hugging Face
Downloadvoices/bf_isabella.safetensors
safetensors · 510 KB · SHA-256 3a3f580cd789…9f08 · Hugging Face
Downloadvoices/jf_gongitsune.safetensors
safetensors · 510 KB · SHA-256 fa63996f4374…4ed4 · Hugging Face
Downloadvoices/jf_tebukuro.safetensors
safetensors · 510 KB · SHA-256 55708217a673…df64 · Hugging Face
Downloadvoices/zf_xiaobei.safetensors
safetensors · 510 KB · SHA-256 07514e1ffb72…b502 · Hugging Face
Downloadvoices/zf_xiaoxiao.safetensors
safetensors · 510 KB · SHA-256 a91ecaa5467b…df21 · Hugging Face
Downloadvoices/zm_yunjian.safetensors
safetensors · 510 KB · SHA-256 dc7446f673d6…df56 · Hugging Face
Downloadvoices/zm_yunyang.safetensors
safetensors · 510 KB · SHA-256 b645dadf7bbe…f126 · Hugging Face
DownloadSource excerpts
2 excerpts
Source excerpts
2 excerpts--- license: apache-2.0 language: - en pipeline_tag: text-to-speech tags: - mlx - kokoro - tts - onyx base_model: hexgrad/Kokoro-82M --- # Onyx-Kokoro-82M A mixed-precision MLX quantization of [Kokoro-82M](https://huggingface.co/hexgrad/Kokoro-82M) (StyleTTS2-derived, Apache-2.0), packaged for **[Onyx](https://github.com/)** — an on-device iOS AI assistant. Used by OnyxTTS, Onyx's own from-scratch Swift + MLX text-to-speech engine, to speak assistant replies in a natural neural voice fully on-device (Apple's AVSpeechSynthesizer is the fallback). ## What this is - **Mixed 4-bit quant** (group size 64): the PL-BERT (ALBERT) text encoder's Linear weights are 4-bit; the entire vocoder path — the prosody predictor (LSTMs + AdaIN convs → pitch/energy) and the iSTFTNet generator — stays fp16. The vocoder directly shapes the waveform and is quant-sensitive, so it is deliberately NOT quantized. ~330 MB fp32 → **148 MB** (`model.safetensors`). - **Voice packs** (`voices/*.safetensors`): 54 voices (af_/am_ American, bf_/bm_ British, plus jf/jm/zf/zm/ ef/em/ff/hf/hm/if/im/pf/pm for other languages). Each is a `[N, 1, 256]` style tensor indexed by phoneme-sequence length. - `config.json` carries the standard Kokoro config plus a `quantization` block (`{group_size: 64, bits: 4}`) so the loader dequantizes the right layers. ## Provenance Weights derived from `prince-canuma/Kokoro-82M` (the mlx-audio mirror of `hexgrad/Kokoro-82M`) via Onyx's `quantize_kokoro.py`. Verified to load + synthesize valid 24 kHz audio end-to-end through OnyxTTS on Apple Silicon. Apache-2.0, same as the base model.
Source context: 106 downloads · 1 likes · Pipeline text-to-speech · Library mlx · Repo wabibito/Onyx-Kokoro-82M
