Chinese tech giant Baidu has made significant strides in document processing technology by open-sourcing Unlimited OCR, a 3 billion parameter Mixture of Experts (MoE) model designed for parsing long documents efficiently. This innovative model addresses a major limitation in traditional OCR systems—how memory and processing time increase exponentially with document length.
Revolutionary Architecture for Long-Document Parsing
The core innovation behind Unlimited OCR lies in its Reference Sliding Window Attention (R-SWA) mechanism. Unlike conventional models that see memory usage and latency grow with output length, R-SWA maintains a constant key-value (KV) cache size. This allows the model to process dozens of document pages in a single forward pass without performance degradation.
This architectural breakthrough is particularly impactful for industries dealing with lengthy documents such as legal contracts, academic papers, and medical records. By keeping the KV cache flat, Unlimited OCR ensures consistent performance regardless of document size—a critical advantage in real-world applications where document lengths vary dramatically.
Performance and Open-Source Impact
Benchmarked on OmniDocBench v1.5, Unlimited OCR achieved a score of 93.23, outperforming the DeepSeek OCR baseline by 6.22 points. The model is released under the permissive MIT license, encouraging widespread adoption and further development in the open-source community.
This release not only underscores Baidu’s growing influence in AI research but also signals a shift toward more scalable and efficient OCR technologies. As document processing demands continue to rise, solutions like Unlimited OCR may become essential tools for businesses and researchers working with large volumes of text-based data.
