[numeri]

Deep seek OCR is an LLM, just one trained/post-trained specifically for OCR.

Exact details of text to image compression ratios are of course extremely dependent on the model architecture, training data, training objectives, etc., so there's probably not too much justification for generalizing to all models