import mlx_lm
import os

# This script converts the heavy 15GB 8-bit model to a 
# 4-bit MLX model optimized for Apple Silicon (M1/M2/M3).
# Size will drop to ~8GB.

model_path = "../model"
output_path = "ankahi_mlx_4bit"

if not os.path.exists(output_path):
    print(f"Converting {model_path} to 4-bit MLX...")
    mlx_lm.convert(
        hf_path=model_path, 
        mlx_path=output_path,
        quantize=True,
        q_group_size=64,
        q_bits=4
    )
    print(f"Success! Optimized model saved to: {output_path}")
else:
    print(f"Optimized model already exists at {output_path}")