Chandra

Datalab, the team behind Marker and Surya, trained Chandra to turn images and PDFs into structured HTML, Markdown, or JSON while preserving layout information. The vision language model handles complex tables, forms, handwriting, and mathematical notation, extracts embedded images with captions, and reads more than 90 languages, averaging 77.8 percent across 43 languages on the project's multilingual benchmark. Installation is pip install chandra-ocr with a choice of vLLM or Hugging Face inference backends, and throughput reaches about 1.44 pages per second on an NVIDIA H100; an NVIDIA GPU is effectively required for practical use. The code is Apache-2.0, but the weights carry a modified OpenRAIL-M license: free for research, personal use, and companies under 2 million dollars in funding or revenue, with commercial licensing sold by Datalab beyond that threshold. The repository has about 11,800 stars, and the Chandra 2 release of March 2026 scores 85.8 on the olmOCR benchmark. Document pipelines adopt it where handwriting, forms, and multilingual scans defeat conventional OCR engines.