Reportedly: Baidu open-sources Wenxin Unlimited OCR 2, enabling arbitrary-length document parsing
Reported by Leiphone · not yet confirmed by the company or a second independent outlet. We update this page when it is.
Baidu released and open-sourced Wenxin Unlimited OCR 2 on October 27, achieving state-of-the-art scores of 84.6 on MPDocBench and 77.1 on MarathonDocBench. The model parsed the entire Harry Potter series, over 3,600 pages, in one pass. Its SR-SWA attention mechanism delivers 7x faster single-step decoding than full attention on hundred-page documents, Leiphone reported.
China context
- Original name
- 文心 Unlimited OCR 2
- Outside China
- Open weights · leiphone.com
- Claims
- Company-reported; not yet independently evaluated
- For builders
- Developers can integrate an open-weights OCR model that handles arbitrary-length documents, eliminating page-limit workarounds. Check the license and model size before deployment.
- For investors
- Baidu's open-source OCR release intensifies competition in document AI, potentially commoditizing long-document parsing and pressuring commercial OCR vendors.
Translated from Chinese. Quotes and facts link to the original sources.
Baidu released and open-sourced Wenxin Unlimited OCR 2, a model for parsing documents of arbitrary length. It achieves state-of-the-art results on MPDocBench (84.6) and MarathonDocBench (77.1), outperforming Qwen 3.8-Max, Gemini-3.8-Flash, GLM-5.3-Flash, and DeepSeek-4.1-Flash. The model uses Sparse Reference Sliding Window Attention (SR-SWA) and Unlimited-Horizon Decoding to reduce computation and memory overhead. In tests, it parsed the entire Harry Potter series (over 3,600 pages) in one pass. SR-SWA improves single-step decoding speed by 7x over full attention and 5.6x over the previous R-SWA on hundred-page documents. The first-generation Unlimited OCR, released June 22, topped GitHub trending and Hugging Face model trending charts, gaining 10,000 GitHub stars in five days.
SR-SWA uses a Page Gate to restrict each query's visual reference to relevant pages and Token Buckets to capture long-range visual dependencies, mimicking how humans copy books by focusing on the current page and only referring back when needed. Unlimited-Horizon Decoding reduces computation and memory for very long documents. Reported speedups: 7x vs full attention, 5.6x vs R-SWA on hundred-page documents.
Developers outside China can now use an open-weights OCR model that handles documents of arbitrary length, removing page limits that previously forced chunking and stitching. This lowers the cost and complexity of processing long reports, books, and archives, and pressures closed OCR APIs to match the capability.
Enterprises with large document archives can reduce processing costs and improve accuracy by using an open-weights model that handles entire documents in one pass, avoiding manual splitting and context loss. The reported speedups may lower inference costs for long documents.
Verify the open-source release by checking the official GitHub repository and Hugging Face model page for weights and license. Independent benchmarks on long-document OCR tasks will confirm the reported SOTA scores. Adoption can be tracked via GitHub stars, forks, and downstream integrations.