How accurate is it?
On a clean picture of a printed table, in testing, the reader read an eight-row invoice with every cell right, header and totals included. On a phone photo taken at an angle, or a table with faint rules and wrapped cells, expect merged or shifted cells and check the sheet against the picture. The Markdown beside the spreadsheet is there for exactly that comparison.
Why is it so much heavier than the other OCR tools?
The fast reader finds lines and reads them, which is enough for text and costs six megabytes. Keeping a table as a table means understanding the page's layout, and the smallest model measured that does it honestly in a browser is 660 MB. That is the weight of rows and columns, and it is downloaded once per browser.
Does the picture leave my device?
No. The model is downloaded to this browser; the picture is read here. The only outside request is for the model's own files, from Hugging Face, pinned to one version and checked against digests this site carries. What that request contains is on the privacy page.
Can it read a table out of a scanned PDF?
A one-page PDF, yes: the page is drawn and read as a picture. For a multi-page scan, use OCR a PDF's Deep mode, which reads every page as Markdown with its tables as rows of cells, or export the page you need with PDF to image and bring it here.
What does Pro cost?
Nothing during the preview. The Pro door shows a price because we want to know whether a reader of this weight is worth paying for; behind it there is no checkout, no card and no account. If a paid tier ever exists, it will be a new page that says so, not a change to this one.