Unlimited OCR 2
An ultra-long document OCR model released by Baidu. It adopts a sparse reference sliding-window attention mechanism and supports unbounded document parsing, achieving state-of-the-art results on ultra-long document benchmarks.
Timeline (last 90 days)
If parsing hundreds of pages in a single request holds up reliably, it will change the current situation in document intelligence pipelines where the chunking strategy after OCR is constrained by a page-count limit.
Baidu has open-sourced the Unlimited OCR 2 model.
Unlimited OCR 2 uses the MIT license.
Unlimited OCR 2 can parse hundreds of pages of documents in a single request.
Unlimited OCR 2 uses the SR-SWA attention mechanism, keeping the decoding cost per step almost constant.
Unlimited OCR 2 scored 77.1 on MarathonDocBench.
SR-SWA's specific implementation details, VRAM usage curves or latency data at different page counts were not disclosed in the original material.
A single benchmark score (such as 77.1) is insufficient to fully assess the real quality of long-document OCR; factors such as layout complexity and scan quality also need to be considered.
Baidu released Unlimited OCR 2 on Hugging Face, with the model page at baidu/Unlimited-OCR-2.
The core innovation of Unlimited OCR 2 is sparse reference sliding window attention (SR-SWA), which compresses text-vision attention interactions into a fixed-size reference set (for example, 512 tokens).
6 more timeline events
All Facts (16)
SR-SWA's specific implementation details, VRAM usage curves or latency data at different page counts were not disclosed in the original material.
50%UnverifiedA single benchmark score (such as 77.1) is insufficient to fully assess the real quality of long-document OCR; factors such as layout complexity and scan quality also need to be considered.
50%UnverifiedUnlimited OCR 2 uses the MIT license.
50%UnverifiedUnlimited OCR 2 can parse hundreds of pages of documents in a single request.
50%UnverifiedUnlimited OCR 2 uses the SR-SWA attention mechanism, keeping the decoding cost per step almost constant.
50%UnverifiedUnlimited OCR 2 scored 77.1 on MarathonDocBench.
50%UnverifiedThe robustness of SR-SWA on real, complex documents, its adaptability to layout diversity, and whether a fixed reference set will lose key details in extremely long documents all require further verification.
50%UnverifiedThe 512-token reference set mentioned officially is an example value, and the configuration space and tuning strategies in actual deployment have not yet been disclosed.
50%UnverifiedThe core innovation of Unlimited OCR 2 is sparse reference sliding window attention (SR-SWA), which compresses text-vision attention interactions into a fixed-size reference set (for example, 512 tokens).
50%UnverifiedSR-SWA makes per-step decoding latency and memory usage barely grow with document length.
50%UnverifiedUnlimited OCR 2's SR-SWA introduces a reference set mechanism on top of the sliding window, enabling text and visual information to interact across windows and avoiding the problem of pure sliding windows forgetting distant information.
50%UnverifiedBaidu says Unlimited OCR 2 achieves SOTA performance on the tens-of-pages-level benchmark MPDocBench and the hundreds-of-pages-level benchmark MarathonDocBench.
50%UnverifiedUnlimited OCR 2's SOTA claim is based on Baidu's official self-testing results, and independent third-party evaluations have not yet appeared.
50%UnverifiedIf parsing hundreds of pages in a single request holds up reliably, it will change the current situation in document intelligence pipelines where the chunking strategy after OCR is constrained by a page-count limit.
50%UnverifiedBaidu has open-sourced the Unlimited OCR 2 model.
50%UnverifiedBaidu released Unlimited OCR 2 on Hugging Face, with the model page at baidu/Unlimited-OCR-2.
50%