AnswersPublished16 sources
Baidu Unlimited OCR 2: How SR-SWA Handles Long Documents
Released on October 27, 2026, Unlimited OCR 2 uses Sparse Reference Sliding Window Attention (SR SWA) to keep its active attention references bounded while parsing multi page documents in one request. Baidu also reports 7.0× faster per step decoding than full attention and 5.6× faster than its earlier Reference Slid...
Published byEdited with GPT-6 LunaImages generated with GPT Image 2
Research answer
Key takeaways
- Released on October 27, 2026, Unlimited OCR 2 uses Sparse Reference Sliding Window Attention (SR SWA) to keep its active attention references bounded while parsing multi page documents in one request.
- Baidu also reports 7.0× faster per step decoding than full attention and 5.6× faster than its earlier Reference Sliding Window Attention in a 100 page setting.
- The code and model weights are available through Baidu’s GitHub project and Hugging Face repository.
AI Prompt
openai.comCreate a landscape editorial hero image for this Studio Global article: What is Baidu’s Unlimited OCR 2, open-sourced as reported on October 28, 2026, and how does it advance the original Unlimited-OCR released i. Article summary: Baidu’s Unlimited OCR 2 is an open-source, end-to-end document-parsing model designed to process extremely long documents in a single request using bounded working attention, rather than a context that grows with every p. Topic tags: general, general web, user generated, academic. Style: premium digital editorial illustration, source-backed research mood, clean composition, high detail, modern web publication hero. Use reference image context only for broad subject, composition, and topical grounding; do not copy the exact image. Avoid: logos, brand marks, copyrighted characters, real person likenesses, fake screenshots, UI text, readable text, watermarks, char
Baidu released Unlimited OCR 2 on October 27, 2026, as an open-source model for parsing long documents in a single request. Its central change is Sparse Reference Sliding Window Attention (SR-SWA), which compresses text–visual attention into a fixed-size set of working references rather than expanding attention across the entire document. Baidu’s reported benchmark and speed results are promising, but they should be treated as company-reported claims, not independent validation. 3
6
2
What changes in Unlimited OCR 2
The first version used Reference Sliding Window Attention (R-SWA); version 2 introduces SR-SWA for longer-horizon parsing. The project describes SR-SWA as reducing attention interactions to a fixed-size working set—for example, 512 tokens—with the goal of keeping per-step decoding latency and memory use nearly constant as document length grows. 6
7
That bounded active reference set is the key idea behind the “unlimited” claim: the model aims to continue parsing without carrying an ever-growing attention context. It does not mean that processing time or every resource stays constant as pages are added, or that the model can parse every document accurately.
The Hugging Face instructions describe multi-page inference through infer_multi: provide page images in order, and the model parses them in one request, with cross-page context available. 5 Baidu also describes the model as using an unlimited-horizon decoding approach.
1
Reported results and demonstration
A news summary of Baidu’s release reports scores of 84.6 on MPDocBench and 77.1 on MarathonDocBench, a benchmark introduced for ultra-long document parsing. Baidu characterizes the results as state of the art; the available reporting does not provide enough detail here to independently compare the model against named competitors. 2
1
Baidu also reports that, in a 100-page setting, per-step decoding was 7.0× faster than full attention and 5.6× faster than R-SWA. These figures describe decoding steps, not the total time to parse a complete document. 1
A circulated demonstration reportedly processed all seven Harry Potter books—more than 3,600 pages—in one run. That illustrates the intended long-document workflow, but the available evidence does not establish an independently audited accuracy score or a complete runtime measurement for the demonstration. 2
Where to find the model
Baidu’s project code is published in the Unlimited-OCR GitHub repository, and the model weights are listed in the Unlimited-OCR-2 Hugging Face repository. 3
6
What the claims do—and don’t—show
The core design claim is specific: SR-SWA limits the active attention reference set, while multi-page inference lets the model process ordered page images in one request. 5
6 That is different from proving constant total memory use or constant end-to-end processing time; those would require measuring the full system, not just per-step decoding or attention.
The benchmark scores, state-of-the-art characterization, and speed comparisons are reported results, not independent evaluations in the sources cited here. They are useful signals of Baidu’s intended performance, but readers should interpret them as claims pending outside reproduction. 1
2
Studio Global AI
Continue your research
This page includes a source-backed answer you can continue inside Studio Global.
People also ask
What is the short answer to "Baidu Unlimited OCR 2: How SR-SWA Handles Long Documents"?
Released on October 27, 2026, Unlimited OCR 2 uses Sparse Reference Sliding Window Attention (SR SWA) to keep its active attention references bounded while parsing multi page documents in one request.
What are the key points to validate first?
Released on October 27, 2026, Unlimited OCR 2 uses Sparse Reference Sliding Window Attention (SR SWA) to keep its active attention references bounded while parsing multi page documents in one request. Baidu also reports 7.0× faster per step decoding than full attention and 5.6× faster than its earlier Reference Sliding Window Attention in a 100 page setting.
What should I do next in practice?
The code and model weights are available through Baidu’s GitHub project and Hugging Face repository.
Sources
huggingface.coREADME.md · baidu/Unlimited-OCR-2 at main
github.comGitHub - baidu/Unlimited-OCR: Unlimited OCR Works: Welcome the Era of One-shot Long-horizon Parsing.
news.supwil.comOctober 27, 2026 · AI & Tech Daily Digest - Swil-News