spokedotso/whisper-large-v3-turbo-4bit
A 4-bit MLX quantization of openai/whisper-large-v3-turbo, hosted by Spoke for on-device speech-to-text on Apple Silicon.
Forked from mlx-community/whisper-large-v3-turbo-4bit so that Spoke serves its own models. The weights are byte-identical to the upstream conversion (same SHA256 checksums).
Parameter count
Hugging Face reports about 0.1B because it counts the packed 4-bit weight tensors (stored as uint32), not logical parameters. The underlying model is Whisper large-v3-turbo with roughly 809M parameters, quantized to 4 bits with a group size of 64.
Usage
Spoke loads this model on-device with mlx-whisper:
pip install mlx-whisper
import mlx_whisper
result = mlx_whisper.transcribe(
"audio.mp3",
path_or_hf_repo="spokedotso/whisper-large-v3-turbo-4bit",
)
print(result["text"])