Unlimited OCR 2 goes open source

Baidu says it has open-sourced Unlimited OCR 2, a parser it claims can handle documents of virtually unlimited length in one go. Its SR-SWA mechanism focuses on the current page while selectively consulting earlier pages; Baidu reports nearly constant per-step decoding latency as documents grow. Baidu Research says it tested the parser on 5,000-page documents and reports SR-SWA was 7.0× faster than Full Attention and 5.6× faster than R-SWA on 100-page documents.