CoolFace
Modelpublic

sanke/trelis-mlx-hinglish-q4

sourceHugging Facecc-by-nc-4.0updated 2mo agoView on Hugging Face
0likes7downloads
Model Card

Trelis Whisper Hinglish (Preview) - MLX Q4

This is a 4-bit MLX quantized version of the Trelis/whisper-hinglish-preview model by Trelis Research.

Conversion Details

The original model (a fine-tune of whisper-large-v3 for English/Hindi code-switching) has been unsharded and converted to the MLX format with 4-bit quantization. This reduces the memory footprint from ~3.1 GB down to ~733 MB, allowing it to run extremely fast on Apple Silicon (M-series) Macs while maintaining high accuracy.

Usage with MLX

You can load and run this model using the mlx-whisper package:

python
import mlx_whisper

text = mlx_whisper.transcribe(
    "audio.wav",
    path_or_hf_repo="sanke/trelis-mlx-hinglish-q4"
)
print(text["text"])

Credits & Links