Unlimited-OCR
lastmod 2026-10-30created 2026-10-301 edits
An OCR model that parses documents of hundreds of pages by carrying context across pages
About
- Baidu released Unlimited-OCR 2, which processes multiple pages in a single request. It uses cross-page context and is designed to keep parsing very long documents within a finite context window
- SR-SWA compresses the attention interaction between text and visual information into a fixed-size reference set, keeping per-step decoding latency and memory usage nearly constant even as documents grow longer
- In its own published evaluation, it scored 84.6 on MPDocBench (dozens of pages) and 77.1 on MarathonDocBench (hundreds of pages), the highest overall score among the compared models
- Provides Transformers for quick testing and a customized SGLang for serving. Published examples for single-page and multi-page input, and for converting PDFs to images for processing
- Shrinking the page window used to split long documents lowers peak memory usage, but the context of blocks that continue across boundaries is cut off, which can degrade cross-page merge quality