Baidu's 'Unlimited OCR' Processes Dozens of Pages in a Single Pass

Baidu researchers have developed 'Unlimited OCR,' a new optical character recognition model that can process dozens of document pages in a single inference pass without increasing memory use or impacting speed based on text length. This is achieved through a redesigned attention mechanism called Reference Sliding Window Attention (R-SWA), which keeps the KV cache (a buffer for processed tokens) constant. The model, built on Deepseek OCR and utilizing a Mixture-of-Experts architecture, achieves 93.92% on the OmniDocBench v1.6 benchmark and significantly reduces error rates even on long documents.