Baidu ships Unlimited OCR 2, bounding text-visual attention to a fixed reference set for unlimited-horizon parsing
Synopsis
Baidu released Unlimited OCR 2, aiming to move document parsing from long documents to unlimited-horizon parsing. Earlier OCR models let text and visual tokens interact across ultra-long sequences, so computation grew with length; the new version uses Sparse Reference Sliding Window Attention (SR-SWA) to compress those interactions onto a fixed-size set of working references, for example 512 tokens, bounding computational complexity. The release also introduces MarathonDocBench, a benchmark for ultra-long document parsing, and the model is served through Transformers, vLLM and SGLang.
Interpretation
Baidu released Unlimited OCR 2, extending the parsing target from long documents to unlimited-horizon parsing, which the release frames as the start of the OCR-3.0 era. The earlier Unlimited-OCR was positioned as pushing Deepseek-OCR one step further; this release states the scope as moving from long-horizon to unlimited-horizon parsing. The release ties unlimited-horizon parsing directly to SR-SWA and ships MarathonDocBench, a benchmark for ultra-long document parsing.
The central change is Sparse Reference Sliding Window Attention, which makes text-visual attention interactions happen over a fixed-size set of working references. The usual route lets text and visual tokens interact pairwise across ultra-long sequences, so the interaction count grows with sequence length; SR-SWA compresses those interactions onto a fixed-size reference set. The release gives 512 tokens as an example reference size and states that this bounds computational complexity.
Perspective
The release addresses researchers and engineering teams parsing ultra-long documents; its stated scope is that SR-SWA compresses text-visual interactions onto a fixed-size reference set, accompanied by MarathonDocBench, a benchmark for ultra-long document parsing.
The release states that SR-SWA bounds computational complexity, but the reference size it gives is only the 512-token example, without describing behavior across different document lengths.