Baidu has released Unlimited OCR 2, an open-source OCR model designed for parsing very long documents such as books, filings, and reports in a single request.
Key points
- Baidu released Unlimited OCR 2 for parsing hundreds of pages.
- The model uses SR-SWA sparse attention for efficient decoding.
- Unlimited OCR 2 achieved first place on multi-page benchmarks.
- The model has 3.34 billion parameters and a 131,072 token max length.
This 3.34 billion parameter model achieves this capability through a core innovation called SR-SWA, a sparse attention scheme that maintains a nearly flat per-step decoding cost.
The model transforms ordered page images into structured text, including text, tables, formulas, figures, headings, and reading order. Unlimited OCR 2 succeeds Unlimited-OCR, released in June 2026, and features a configuration that maintains a DeepSeek-OCR style layout. It employs a SAM ViT-B plus CLIP-L encoder feeding a 12-layer mixture-of-experts decoder.
By the numbers
- parameter model size
- 3.34B
- max position length
- 131,072
- MarathonDocBench overall score
- 77.1
SR-SWA works by compressing text-visual attention into a fixed set of working references, addressing the quadratic growth in full attention cost with sequence length. Baidu tested Unlimited OCR 2 on an NVIDIA A800-80GB GPU using BF16 precision, with weights totaling approximately 6.7 GB. The model achieved first place on multi-page benchmarks tested by Baidu, scoring 77.1 overall on MarathonDocBench.
In performance tests, Unlimited OCR 2 scored 77.1 overall on MarathonDocBench, significantly outperforming the next model, Kimi-K3, which scored 33.7. Its lowest sub-score on MarathonDocBench was 48.2 for heading hierarchy (TEDS). The model supports a maximum position length of 131,072 tokens and runs on a single CUDA GPU in BF16.
Sources and references
Original report
Baidu Open-Sources Unlimited OCR 2 for Hundred-Page Document Parsing
MarkTechPost
Read on MarkTechPost (opens in a new tab)Linked in the original report
- Modelhuggingface.co/baidu/Unlimited-OCR-2 (opens in a new tab)
- Codegithub.com/baidu/Unlimited-OCR (opens in a new tab)
- Modelhuggingface.co/baidu/Unlimited-OCR (opens in a new tab)
- Codegithub.com/deepseek-ai/DeepSeek-OCR (opens in a new tab)
- Modelhuggingface.co/baidu/Unlimited-OCR-2/blob/main/config.… (opens in a new tab)
- Paperarxiv.org/abs/2510.18234 (opens in a new tab)
Ref 00UXXYBI
Cite this
APA
Interestana AI Editorial. (2026, October 27). Baidu Releases Unlimited OCR 2 for Long Document Parsing. Interestana. https://interestana.com/articles/baidu-open-sources-unlimited-ocr-2-for-hundred-page-document-parsing-00uxxybi
MLA
“Baidu Releases Unlimited OCR 2 for Long Document Parsing.” Interestana, 27 Oct. 2026, interestana.com/articles/baidu-open-sources-unlimited-ocr-2-for-hundred-page-document-parsing-00uxxybi.
Chicago
Interestana AI Editorial. “Baidu Releases Unlimited OCR 2 for Long Document Parsing.” Interestana, October 27, 2026. https://interestana.com/articles/baidu-open-sources-unlimited-ocr-2-for-hundred-page-document-parsing-00uxxybi.
BibTeX
@misc{interestana_00uxxybi,
author = {{Interestana AI Editorial}},
title = {{Baidu Releases Unlimited OCR 2 for Long Document Parsing}},
howpublished = {Interestana},
year = {2026},
month = oct,
url = {https://interestana.com/articles/baidu-open-sources-unlimited-ocr-2-for-hundred-page-document-parsing-00uxxybi},
note = {Summary of reporting by MarkTechPost, https://www.marktechpost.com/2026/10/26/baidu-unlimited-ocr-2-open-source-long-document-ocr/}
}Credit the source
Original reporting: MarkTechPost, “Baidu Open-Sources Unlimited OCR 2 for Hundred-Page Document Parsing”, https://www.marktechpost.com/2026/10/26/baidu-unlimited-ocr-2-open-source-long-document-ocr/(opens in a new tab)
Follow this story
Get Computing & Chips in your inbox — new stories in 40 words or fewer, once a week.
Subscribe