{"version":1,"type":"story","url":"https://digestai.news/story/baidu-releases-unlimited-ocr-2-with-unlimited-horizon-parsing-via-sr-s","json":"https://digestai.news/story/baidu-releases-unlimited-ocr-2-with-unlimited-horizon-parsing-via-sr-s.json","markdown":"https://digestai.news/story/baidu-releases-unlimited-ocr-2-with-unlimited-horizon-parsing-via-sr-s.md","slug":"baidu-releases-unlimited-ocr-2-with-unlimited-horizon-parsing-via-sr-s","headline":"Baidu releases Unlimited-OCR-2 with unlimited-horizon parsing via SR-SWA","summary":"Baidu released Unlimited-OCR-2 on October 27, 2026, introducing Sparse Reference Sliding Window Attention (SR-SWA) that compresses text-visual attention interactions to a fixed-size set of working references such as 512 tokens. This bounds computational complexity and keeps per-step decoding latency and memory footprint nearly constant regardless of document length. The model achieves state-of-the-art performance on both dozens-of-pages benchmarks like MPDocBench and extreme hundred-page benchmarks like MarathonDocBench. Unlimited-OCR-2 supports Transformers for quick tests and SGLang for serving, with tested dependencies including Python 3.10, torch 2.13.0, and transformers 4.52.1 on an NVIDIA A800-80GB. The release includes multi-page inference that parses all pages in one request using cross-page context, and single-page inference with dynamic tiling. Detailed SGLang server configuration instructions are provided with environment variables for memory optimization. The model is available on Hugging Face, ModelScope, and Baidu Cloud, with a demo on Hugging Face Spaces. The accompanying arXiv paper (2606.23050) lists authors including Youyang Yin, Huanhuan Liu, and others from Baidu.","keyPoints":["Baidu releases Unlimited-OCR-2 with SR-SWA for unlimited-horizon parsing","Achieves SOTA on MPDocBench and MarathonDocBench benchmarks","Available on Hugging Face, ModelScope, Baidu Cloud with SGLang serving support"],"whyItMatters":"Unlimited-OCR-2 eliminates quadratic attention scaling, enabling parsing of entire books in a single request without memory bottlenecks, advancing document AI for enterprise and research.","category":{"slug":"models","name":"Generative AI & Models","url":"https://digestai.news/category/models"},"entities":{"companies":["Baidu","Hugging Face","ModelScope","Baidu Cloud","SGLang","vLLM"],"models":["Unlimited-OCR-2","Deepseek-OCR","Deepseek-OCR-2","PaddleOCR"],"people":["Youyang Yin","Huanhuan Liu","Qunyi Xie","Chaorun Liu","Shiqi Yang","Shaohua Wang"]},"firstPublishedAt":"2026-10-27T08:54:41Z","updatedAt":"2026-10-27T08:54:41Z","sourceCount":1,"hasPrimarySource":true,"sources":[{"outlet":"huggingface.co","title":"baidu/Unlimited-OCR-2 · Hugging Face","url":"https://huggingface.co/baidu/Unlimited-OCR-2","publishedAt":"2026-10-27T08:54:41Z","type":"primary","primary":true,"lead":true}],"sourceNotes":null,"discussions":[{"site":"Reddit","url":"https://www.reddit.com/r/LocalLLaMA/comments/1xgnbtq/baiduunlimitedocr2_hugging_face/","points":null}],"thread":null,"cite":{"text":"Digest AI, \"Baidu releases Unlimited-OCR-2 with unlimited-horizon parsing via SR-SWA\", 27 October 2026, https://digestai.news/story/baidu-releases-unlimited-ocr-2-with-unlimited-horizon-parsing-via-sr-s","publisher":"Digest AI","title":"Baidu releases Unlimited-OCR-2 with unlimited-horizon parsing via SR-SWA","datePublished":"2026-10-27T08:54:41Z","url":"https://digestai.news/story/baidu-releases-unlimited-ocr-2-with-unlimited-horizon-parsing-via-sr-s"},"generatedBy":"Written by Digest AI's editorial model from the linked sources; the sources are the record.","license":"Headlines, digests and key points are written by Digest AI and may be quoted with a link to the story page. Linked articles belong to their publishers. Terms: https://digestai.news/terms#reuse"}