Baidu Open-Sources Unlimited OCR 2 for Long Documents
On October 27, 2026, Baidu open-sourced Unlimited OCR 2 under MIT on Hugging Face and GitHub—about 3.3B parameters. Sparse Reference Sliding Window Attention (SR-SWA) compresses ultra-long text-visual attention into a fixed-size set of working references so per-step latency stays nearly constant as documents grow. Baidu reports SOTA on MPDocBench and MarathonDocBench; those scores are vendor-reported. Multi-page parsing can run in one request via infer_multi, with Transformers for quick tests and SGLang for serving.
