Kokoro-82M — bf16 (MLX)

Kokoro-82M converted to bfloat16 for MLX inference on Apple Silicon.

Converted from mlx-community/Kokoro-82M-bf16 revision a71e4d38b236d968966a2002c4c895dbd12b1c3c — whose safetensors are stored as float32 despite the repo name. This conversion halves the download (156 MB vs 312 MB) with no quality change: the model's architecture (ALBERT encoder + BiLSTM + ISTFTNet vocoder) runs identically at bf16 precision on Apple Silicon's native bf16 hardware.

Files

File	Size	SHA-256
`model.safetensors`	163,588,165 bytes (156.0 MB)	`235a936cbf762c07625543eec9f76af08bb401151534e692307ecc3ce80c5818`
`config.json`	2,351 bytes	`5abb01e2403b072bf03d04fde160443e209d7a0dad49a423be15196b9b43c17f`

Voice pack (voices-v1.0.bin, ~28 MB) is unchanged from the original and hosted at thewh1teagle/kokoro-onnx.

Provenance

Source: mlx-community/Kokoro-82M-bf16 @ a71e4d38
Original model: hexgrad/Kokoro-82M (Apache 2.0)
Conversion: F32 → bfloat16 via mlx.core (no quantization, no pruning)
Every original F32 tensor verified as float32 before conversion
Original digests cross-checked against the pinned manifest

Usage with M1K3

These weights are used by M1K3's Kokoro neural voice (M1K3Kokoro target). The app loads them via MLX.loadArrays — dtype is handled transparently by mlx-swift.

Downloads last month: 123

Safetensors

Model size

81.8M params

Tensor type

BF16

MLX

Hardware compatibility

Quantized

Model tree for round-tower/Kokoro-82M-bf16

Base model

yl4579/StyleTTS2-LJSpeech

Finetuned

hexgrad/Kokoro-82M

Finetuned

(60)

this model