Unlimited-OCR BF16 MLX

MLX conversion of baidu/Unlimited-OCR.

from mlx_vlm import load
from mlx_vlm.generate import generate

model, processor = load("mikoy92/Unlimited-OCR-bf16-mlx")
result = generate(
    model,
    processor,
    "<image>\ndocument parsing.",
    image="your_image.jpg",
    max_tokens=512,
    temperature=0.0,
)
print(result.text)

CLI:

python -m mlx_vlm generate \
  --model mikoy92/Unlimited-OCR-bf16-mlx \
  --image your_image.jpg \
  --prompt "document parsing." \
  --temp 0

This repo stores MLX-layout weights in safetensors (format=mlx). It uses the existing deepseekocr MLX implementation because Unlimited-OCR shares that SAM + CLIP-L + DeepSeekV2 OCR architecture shape with different checkpoint dimensions.