Most OCR stacks force a tradeoff: clean text on English print, but tables collapse, equations turn to gibberish, and non-Latin scripts fall apart. Chandra treats the page as a single vision-language model rather than a detect-then-recognize pipeline, so structure and content come out together — a 90-language model that keeps the table a table and the equation an equation.
What Sets It Apart
- Layout is preserved, not reconstructed. Output lands directly as Markdown, HTML, or JSON with the original structure intact, skipping the brittle post-processing most OCR tools need to rebuild tables and headings.
- Strong where OCR usually breaks. It handles handwriting, math, charts, forms, and even chemistry diagrams, posting an 85.8 olmOCR benchmark score and a 72.7% average across 90 languages — well ahead of Gemini 2.5 Flash's 60.8% on the same multilingual set.
- Two deployment modes from one model. Run it locally through HuggingFace for privacy and tinkering, or point it at a vLLM server for throughput (~1.44 pages/sec on a single H100).
Who It's For
Great fit if you process multilingual, table-heavy, or scientific documents and want structured output without stitching together a detection-plus-recognition pipeline. It comes from the team behind Surya and Marker, so it slots naturally into existing document workflows. Look elsewhere if you need permissive licensing for a large commercial product: the code is Apache 2.0, but the model weights ship under a modified OpenRAIL-M that is free only for research, personal use, and smaller companies — bigger shops are steered toward Datalab's paid platform.
