Most OCR tools give you a wall of text: the words are there, but the tables are flattened, the formulas are broken and the headings are gone. Unlimited OCR 2, released by Baidu under the MIT license, takes a different approach: you give it a photo or a scan of a page and it returns Markdown, keeping text, tables, formulas and layout.
Unlimited OCR 2Baidu · MITTurns photos and scans of documents into clean Markdown: text, tables, formulas and layout, page after page.Try it online →What Unlimited OCR 2 is good at
- Tables: rows and columns come back as Markdown tables you can paste into a document or a spreadsheet.
- Formulas: maths is kept as formulas instead of becoming random characters.
- Layout: headings, lists and paragraphs keep their structure, so the output is ready to edit.
- Photos, not only clean scans: a picture of a page taken with a phone works.
It is a 3.3B-parameter vision-language model. That is small for a model of this kind, but it still wants a GPU with about 24 GB of memory to run at full precision (BF16), which most laptops don't have.
How to try it online
- Open the Unlimited OCR 2 page on LiveWeights and sign in (email code or Google).
- Upload a photo or scan of one page (up to 6 MB).
- Press run. You get the Markdown back, ready to copy.
The first run is free, and new accounts also get welcome credits for the other models. If the model has not been used for a while, the first run starts a GPU first (a "cold start" of a couple of minutes); the following runs are much faster.
License: can I use it commercially?
Yes. Unlimited OCR 2 is released under the MIT license, which allows commercial use. The weights are on Hugging Face and the code on GitHub: both are linked from the model page, together with the earlier v1 release.
Tips for better results
- One page per image: crop away the desk and the neighbouring page.
- Shoot straight from above, in good light, without shadows across the text.
- Check numbers in tables before reusing them: OCR is very good, but not perfect.
Other vision models you can try
If you need more than text, WeVisOne-VL 2B by Tencent removes backgrounds, detects objects and reads text in photos, and Groot 19B lets you chat about any image you attach.