Unlimited OCR

I used Unlimited OCR 2 as the ingest step of a “questions over my own documents” assistant, feeding WeKnora. It doesn’t run on a Mac: the code calls .cuda() and relies on Triton kernels, so I used the Hugging Face demo Space. Anonymous use there allows about two pages a day, and the multi-page endpoint asks for more GPU time than the free tier allows. That left the headline claim, long PDFs in one pass, untested.

On a photo of a supermarket receipt, one page took about 100 seconds. It got about 97% of the English text and numbers right, but it dropped one price, read the “BOGA superfoods” logo as “BOOK / SURVEY BOOKS”, and got most of the Arabic wrong. The output is its own tagged block format with bounding boxes, which I had to convert to Markdown myself. The missing price mattered later: WeKnora’s agent confidently answered that the item cost 1 AED, taking the quantity as the price.

On enterprise fit: I didn’t score it, because I only used the public demo. The model is MIT-licensed and could be self-hosted on NVIDIA GPUs, but I didn’t try that.