What is available
Baidu has published Unlimited-OCR-2, a model for turning page images into editable text while using context across pages. That matters for long manuals and reports where a table or paragraph continues onto the next page. The card dates version 2 to October 27, but the weight files have October 26 commit dates. This is a newly discovered catch-up; we did not establish the first public release hour. [1][3]
The public listing contains two weight files totaling about 6.69 GB, with about 3.34 billion parameters listed; the whole repository displays 6.87 GB. The card labels the model MIT, and the linked repository license requires retaining its copyright and permission notice. That gives experienced technicians concrete files, documentation, and license terms to inspect before considering an evaluation. [1][3][4]
Longer documents still need checks
Baidu’s Sparse Reference Sliding Window Attention, or SR-SWA, limits the working references used during decoding. The card’s 512-token example describes that working set, not the model’s whole context or unlimited computing capacity. The authors report results on their MarathonDocBench: 100 documents spanning 103–300 pages. Those are source-reported evaluations, not tests by Tinker & Tokens. [1][2]
The paper also acknowledges limited training-data diversity and difficulty preserving global document information, including heading hierarchies. Its near-constant per-step decoding measurements do not mean a longer document takes the same total time. The implementation instructions use custom code, Transformers, or a customized SGLang setup; the authors document an NVIDIA A800 80 GB test machine. That is their setup, not an established minimum requirement. [1][2]
Our take
For experienced technicians, the useful question is whether the output preserves a document’s meaning. Review the custom code before installing it. Start with fictional documents or material approved for evaluation. Compare numbers, table headers, reading order, and page boundaries against the original. A clean-looking transcription can still attach a value to the wrong row or turn a subsection into a main heading.
Keep file size and runtime memory separate. Downloaded weights are only part of the memory bill; page images, temporary work, and serving settings also matter. Decide what an acceptable result looks like before comparing tools, and count the time spent checking and correcting output. A model that reads more pages is useful only if the resulting document remains trustworthy enough for the intended job.