wtf is unlimited-ocr?

Unlimited-OCR Works, built by Baidu, is a tool that reads text from images and multi-page PDFs and converts it into structured, editable content. Instead of just spitting out raw text, it preserves the layout, headings, tables, lists, reading order, so the output actually looks like the original document, not a jumbled wall of words. You feed it an image or a PDF, and the model processes the pages, recognizing text and reconstructing the document's structure. It handles single images and multi-page documents, converting PDF pages to images first, then parsing each one. The project includes settings to prevent the model from repeating itself, which is a common problem with long documents. You can run it locally on an NVIDIA GPU, deploy it as a server using tools like vLLM or SGLang, or try a live demo hosted on Hugging Face. This is useful for anyone who needs to digitize large documents, scanned contracts, research papers, financial reports, textbooks, or historical archives. For example, a startup building a document-search product could use it to extract clean, structured text from thousands of PDFs. A legal team could convert scanned case files into searchable data. The emphasis on "long-horizon parsing" means it's designed to handle lengthy documents that would overwhelm typical OCR tools. The project builds on prior work from DeepSeek-OCR and PaddleOCR, pushing those ideas further. It supports multiple deployment options, from a simple Python script to production-ready serving with vLLM or SGLang, including Docker images for easier setup. The README doesn't dive deep into accuracy benchmarks or limitations, so you'd need to test it on your own documents to see how well it performs on your specific use case.