Released June 22, 2026: Baidu Unlimited OCR transcribes 40+ pages in a single 32,000-token forward pass. Weights public under CC-BY 4.0. New end-to-end SOTA on OmniDocBench v1.5 and v1.6.
R-SWA differs from older sliding-window approaches by never subjecting visual reference tokens to recurrent state updates, preventing image degradation over long outputs. Hugging Face demo available. Practical for RAG pipelines processing multi-page PDFs without chunking.