Baidu releases Unlimited OCR 2 on Hugging Face, claiming unlimited-length document parsing

On October 27, 2026, Baidu announced the open-source release of Unlimited OCR 2. The day unfolded across several reports that together fill in the technical picture.

At 10:41 a.m. (UTC+8), Baidu's official @Baidu_Inc account posted on X that Unlimited OCR 2 is open source and introduced SR-SWA, a mechanism that condenses text-visual attention interactions into a fixed-size set of working references. The post said the mechanism keeps long-document attention computation in check, without giving concrete figures. At 1:26 p.m., MarkTechPost reported the model is a 3B OCR model released under the MIT license, capable of parsing hundred-page documents in a single request, with SR-SWA keeping per-step decoding cost nearly flat and a score of 77.1 on MarathonDocBench.

At 2:21 p.m., Leiphone reported that Baidu released and open-sourced the model on October 27, claiming the first parsing of documents of any page count, pushing the OCR ceiling from tens of pages to "unlimited" and reaching end-to-end SOTA. It listed public benchmark composite scores of 84.6 on MPDocBench and 77.1 on MarathonDocBench, and said the model was tested on the full Harry Potter series, seven books totaling more than 3,600 pages parsed in one go while maintaining accuracy and efficiency. The mechanism, Sparse Reference Sliding Window Attention (SR-SWA), was described as borrowing from how humans copy manuscripts: focusing on the current page and checking earlier context only when needed.

At 4:35 p.m., QbitAI reported a new open-source OCR model that compresses KV Cache to a near-constant footprint, cutting cache usage by 96% so that page count is no longer a limit; the report did not name the model or list its parameters and benchmark scores. At 4:54 p.m., the Reddit community LocalLLaMA surfaced the Hugging Face release notes for Unlimited OCR 2, which state that SR-SWA compresses text-visual attention interactions from ultra-long sequences down to a fixed-size set of working references of around 512 tokens, bounding computational complexity so per-step decoding latency and memory footprint stay nearly constant regardless of document length, and framing the release as moving from long-horizon to unlimited-horizon parsing.

On October 28 at 2:29 p.m., OSChina carried the same claims: the first parsing of documents of any page count, the ceiling pushed to "unlimited" at SOTA level, and sustained accuracy and efficiency when parsing a book collection of more than 3,600 pages in one pass. No newer developments have been reported since.

Note that the KV Cache reduction figure (96%) appears in the QbitAI report and not in the release notes, and that the 3,600-page figure refers to the Harry Potter series as described by Chinese media.