Chandra-OCR (github.com)

Chandra-OCR has been announced as a state-of-the-art Optical Character Recognition (OCR) model specifically designed to handle complex documents. It can accurately process handwriting, tables, math equations, and messy forms, which are typically challenging for traditional OCR systems. Users can run Chandra-OCR via HuggingFace Transformers for local inference or deploy it using a vLLM server for enhanced production throughput. The model also outputs results in structured formats like Markdown, HTML, or JSON, complete with layout metadata and bounding box coordinates for every text block and image. This development is significant for the AI/ML community as it expands the capabilities of document processing technologies, especially in sectors requiring the extraction of information from diverse formats, such as finance and education. Chandra-OCR supports over 40 languages and handles intricate layouts, making it an invaluable tool for researchers and businesses alike. By effectively reconstructing forms, preserving tabular structures, and rendering math equations as LaTeX, this model opens new avenues for automating workflows that involve complex documentation tasks. Its open-source nature under a modified OpenRAIL-M license further encourages innovation and adoption in the field.