Unlimited-OCR-2 tested: parsing 30-page PDFs in one request without breaking tables

KhuyenTran16 · x · 2026-11-10

OpenTeams senior DevRel Khuyen Tran hands-on tested Unlimited-OCR-2, Baidu's open-source document parsing model, targeting a classic OCR pain point.

Problem: most OCR models read one page at a time, so tables spanning two pages of a long PDF come back as two separate tables.
Solution: built on DeepSeek-OCR and sized like DeepSeek-OCR-2, Unlimited-OCR-2 reads many pages in a single request (tested with 2–30 page requests), so cross-page tables merge back into one table and paragraphs stay continuous.
How: instead of rereading every page while generating each token (too slow), it splits each page into small bands, summarizes them, then generates across page breaks.
Test scripts are open-sourced on GitHub; the author argues benchmark scores don't always reflect real-world documents, hence testing on an actual long PDF report.

More from coding & agent