AI BriefingPickUnlimited-OCR-2: OCR model removes document length limits with SR-SWA
baidu/Unlimited-OCR-2
About the project
It parses documents of hundreds of pages from start to finish within a single context window. Unlike existing methods, it is designed so that memory and computational costs do not increase exponentially with document length. Unlimited-OCR-2, released by Baidu, targets this 'Unlimited-Horizon' processing.
The core is the Sparse Reference Sliding Window Attention (SR-SWA) technique. It compresses text-visual attention interactions of ultra-long sequences into a fixed-size working reference set of 512 tokens. This keeps per-step decoding latency and memory usage nearly constant regardless of document length.
It delivers top-tier performance on benchmarks of dozens of pages such as MPDocBench, as well as extreme hundred-page-scale benchmarks like MarathonDocBench. It provides basic inference via the Transformers library, along with SGLang support for server deployment. It is released under the MIT license and supports multilingual processing.
It is suitable for developers who need large-scale document digitization or information extraction from long documents. It is particularly strong in multi-page inference tasks that require leveraging cross-page context. It ensures compatibility with major inference frameworks such as vLLM and SGLang, making it easy to apply in production environments.
baidu/Unlimited-OCR-2
The original page has no description.
image-text-to-text
This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.