If you pick an OCR library by glancing at the license badge on its GitHub page, these two will catch you out.
Marker and Surya (both from Datalab) show Apache-2.0 in GitHub's sidebar, and that is accurate ā for the code. The model weights are under a separate licence: a modified AI Pubs Open Rail-M, free for research, personal use, and startups under $5M in funding or revenue. Past that line you need a commercial licence from Datalab. Both projects say so plainly in a "Commercial usage" section of their README; it just isn't what the badge reports, and the badge is what most roundups copy.
The distinction matters because the badge is doing the opposite of its job here. Apache 2.0 on the repo tells you that you may use, modify and redistribute the code. It says nothing about your right to run the weights in a product ā and without the weights the code does not do OCR.
One thing worth flagging if you are reading older comparisons: Marker used to be GPL-3.0 with a $2M cap. It has since relicensed the code to Apache 2.0 and both projects now sit at $5M. So a post from six months ago can be wrong in your favour or against it, in either direction. Check the README and the LICENSE file, not a table someone else compiled.
If you need something with no revenue test at all, Docling is MIT, and Tesseract, olmOCR, PaddleOCR and RapidOCR are all Apache 2.0 end to end.
Fuller comparison of the OCR options, hosted and self-hosted: https://toolfreebie.com/free-ocr-api-pdf/
