Baidu Open-Sources Unlimited OCR 2: Processes Hundreds of Pages in a Single Request, Decoding Cost Does Not Increase with Length
Baidu recently open-sourced the 3B-parameter long-document OCR model Unlimited OCR 2 under the permissive MIT license, supporting direct parsing of documents up to hundreds of pages in a single request. Its self-developed SR-SWA attention mechanism keeps per-step decoding costs nearly constant, and the model scored 77.1 on the long-document OCR benchmark MarathonDocBench.
OCR (Optical Character Recognition) converts text and layout information from images and scanned PDFs into editable structured content. Baidu has recently open-sourced Unlimited OCR 2, a model tailored for long-document scenarios, with 3 billion parameters and the MIT open-source license, with no extra barriers for personal or commercial use.
In daily work, when dealing with hundreds of pages of scanned contracts, academic papers, or archived files, most general OCR tools require splitting pages in advance for sequential recognition. Beyond the cumbersome process, content such as cross-page tables, continuous headers and footers, and page-end annotations are prone to errors when results are combined. Even models supporting long inputs usually have slower recognition speeds and linearly rising computing costs as the number of pages increases, making high-frequency call costs unaffordable for most teams.
Unlimited OCR 2 supports direct parsing of documents with hundreds of pages in a single request, eliminating the need for pre-pagination. It is equipped with Baidu’s self-developed SR-SWA attention mechanism. According to official data, this structure keeps the model’s per-step decoding cost nearly flat, without a noticeable increase as the input document length grows. On MarathonDocBench, a benchmark specifically designed for long-document OCR, the model achieved a comprehensive score of 77.1.
For traditional industries with needs for bulk archival digitization, or teams developing document knowledge bases and AI reading tools, this open-source model can save a significant amount of extra work on long-document preprocessing and computing resource scheduling. The MIT license has no copyleft requirements; as long as users retain the original copyright notice, they are free to modify and redistribute the model.
The news comes from a report by tech media MarkTechPost dated October 26, 2026, authored by Asif Razzaq.
发布时间: 2026-10-27 13:26