Baidu has announced the release of Unlimited-OCR 2, a new model designed to handle document parsing across extremely long sequences. According to the company’s release notes on GitHub, the update introduces a mechanism called Sparse Reference Sliding Window Attention (SR-SWA) to address the computational challenges typically associated with large-scale document processing.
Technical Advancements in Document Parsing
The core technical development in Unlimited-OCR 2 is the implementation of SR-SWA. This mechanism compresses text-visual attention interactions from ultra-long sequences into a fixed-size set of working references, such as 512 tokens.
Baidu claims this approach bounds computational complexity, which maintains nearly constant per-step decoding latency and memory footprint, regardless of the total length of the document. By utilizing this method, the company states the model moves from “long-horizon” to “unlimited-horizon” parsing, effectively removing the quadratic attention scaling and memory bottlenecks that often occur as document length increases.
Benchmark Performance
Baidu reports that the model has achieved state-of-the-art (SOTA) performance in ultra-long document parsing. The company cites high performance across several benchmarks, including MPDocBench, which covers documents spanning dozens of pages, and MarathonDocBench, which tests the model’s capabilities on extreme hundred-page documents.
Early Reception and Comparative Performance
Following the release on October 27, 2026, the model has drawn mixed reactions from the developer community on platforms such as Reddit. While the technical specifications suggest significant improvements in processing efficiency for large files, some users have raised questions regarding its accuracy in specific use cases.
In unverified discussions on Reddit, one user reported that the model performed poorly on French legal documents when compared to LightOnOCR3-0.8B. Other users have questioned how the model compares to existing solutions like PaddleOCR-VL-1.5 and MinerU, with some suggesting that MinerU may hold a performance advantage in certain metrics.
While the technical claims regarding the SR-SWA mechanism and the mitigation of memory bottlenecks are central to the release, independent verification of the SOTA claims on the MarathonDocBench and MPDocBench benchmarks has not yet been established.