Baidu Releases Unlimited OCR, a 3B Model That Keeps the KV Cache Flat for Long-Document Parsing

An open-source OCR model addresses a fundamental problem: as document-parsing models generate more text, their memory requirements grow, slowing down processing. This new model keeps memory constant by using a modified attention mechanism called Reference Sliding Window Attention, which stores only visual tokens and the most recent output tokens rather than all previous outputs. The model achieves this while maintaining high accuracy on document parsing benchmarks and requiring only 500 million active parameters during inference. This design enables practical processing of entire documents, including dozens of pages, in a single pass rather than page-by-page processing.