Baidu Open-sources 3B Model Unlimited OCR: Star Count Exceeds 10,000 in 5 Days, Setting a New Record for Long Document Parsing
Baidu open-sources a 3B-parameter end-to-end OCR model called Unlimited OCR, specifically designed for long documents such as books and papers. The project exceeded 10,000 GitHub stars within 5 days and topped four trending lists. Technically, the model activates approximately 570M parameters, and it innovatively introduces the Reference Sliding Window Attention mechanism, breaking the limitation of page-by-page stitching, supporting continuous parsing of dozens of pages at once, and significantly improving the efficiency of processing long documents.