Baidu Open-Sources Unlimited OCR 2, Parsing Hundred-Page Documents in a Single Request
The 3B model's SR-SWA attention keeps decoding cost near-constant regardless of document length, claiming SOTA on MPDocBench and MarathonDocBench.
Baidu released Unlimited OCR 2, an open-source 3-billion-parameter document parsing model that parses books, filings and reports of hundreds of pages in a single request while maintaining cross-page context. The key technique is Sparse Reference Sliding Window Attention (SR-SWA), which compresses text-visual attention interactions into a fixed-size set of working references, so per-step decoding latency and memory stay nearly constant regardless of document length, eliminating quadratic scaling. Multi-page mode uses 273 image tokens per page.
The company claims state-of-the-art results on MPDocBench and on MarathonDocBench, a new benchmark of very long documents it released alongside the model on 27 October on GitHub and Hugging Face. The release closes a gap left by the week's wave of smaller parsers, which are optimised for short-to-medium documents: thick documents such as annual reports and audited financial statements previously required chunking and stitching, a hidden source of split tables and broken reading order. Benchmark figures are vendor-reported via Baidu's own benchmarks.