Baidu Unveils Unlimited-OCR: Constant KV Cache Delivers SOTA Performance on Long Documents

Baidu has launched Unlimited-OCR, addressing a key issue in long-document transcription. The model employs Reference Sliding Window Attention (R-SWA) to compress KV cache from linear to constant growth, achieving state-of-the-art results on the OmniDocBench benchmark. Built on DeepSeek-OCR's DeepEncoder architecture, Unlimited-OCR scored 93.92% on OmniDocBench v1.6, surpassing DeepSeek-OCR's 87.01%. This improvement highlights significant advancements in transcription accuracy and efficiency.