Because this is a vision-language OCR model, mlx-vlm does not aggressively quantize every multimodal tensor; the effective bits-per-weight can be higher than exactly 4-bit.

Usage

bash

pip install -U mlx-vlm

mlx_vlm.generate \
  --model mikoy92/Unlimited-OCR-4bit-mlx \
  --image /path/to/image.png \
  --prompt "Extract all readable text from this image." \
  --max-tokens 512 \
  --temperature 0

Validation

Before upload, this checkpoint was loaded locally with mlx_vlm.generate and produced OCR text/table output on a document-image smoke test.