Baidu open-sources Wenxin Unlimited OCR 2 model that can parse 3,600+ pages at once

Baidu has released and open-sourced its Wenxin Unlimited OCR 2 model. It lifts document parsing limits from dozens of pages to "unlimited" — it can process all 7 Harry Potter books, over 3,600 pages, in one go.

The model can identify and split layout elements on its own, and merge text and tables across pages. With the same computing power, it handles more documents, faster and at lower cost, which works well for digitizing long documents.

The model scored 84.6 on MPDocBench, a benchmark for multi-page document parsing, and 77.1 on MarathonDocBench, which tests how well it handles 100+ page documents. Both results set the best end-to-end scores so far.