Baidu launches open-source ocr model for efficient hundred-page document processing

Baidu has introduced Unlimited OCR 2, an open-source OCR model designed to process extensive documents with high efficiency. This 3B vision-language model can parse books, filings, and reports spanning hundreds of pages in a single request. The model's core innovation is SR-SWA, a sparse attention mechanism that maintains nearly constant decoding costs as document length increases. It is released under the MIT license and available on GitHub. Unlimited OCR 2 builds upon its predecessor, Unlimited-OCR, which was launched in June 2026. The model employs a SAM ViT-B plus CLIP-L encoder paired with a 12-layer mixture-of-experts decoder. It scores 84.6 on the MPDocBench benchmark and 77.1 on MarathonDocBench, outperforming other models like Qwen3.8-Max and Gemini-3.8-Flash. Baidu highlights the model's ability to handle large documents while maintaining performance across different benchmarks.