Chandra OCR 2: 4B-Parameter Model Scores 85.9% on olmOCR Benchmark — Smaller and More Accurate Than 9B Predecessor
GitHub·medium signal
Datalab's Chandra OCR 2 handles complex tables (merged cells, colspan/rowspan), handwriting (cursive and messy print), forms (checkboxes, radio buttons), and math equations as LaTeX. The 4B model beats the original 9B Chandra across every category. Outputs structured HTML/Markdown/JSON preserving layout. Trending at 9K stars with +191 today. Models on HuggingFace.