MLX conversions of microsoft/Phi-4-multimodal-instruct for native vision-language inference on Apple Silicon (M1-M4). 4-bit, 8-bit, and bf16 variants.