Skip to main content
Back

Datalab's Chandra OCR 2 Beats GPT-4o on Documents at Half the Size

Chandra OCR 2 is a 4B open-source model that tops the olmOCR benchmark at 85.9%, beating GPT-4o and Gemini Flash on structured document extraction across 90+ languages.

  • Chandra OCR 2 is a 4B open-source model from Datalab that scores 85.9% on the olmOCR benchmark, topping GPT-4o, Gemini Flash, and Mistral OCR.
  • Model size dropped from 9B to 4B parameters vs. Chandra 1, while improving accuracy across every benchmark category.
  • Supports 90+ languages, with major gains in South Asian scripts (Malayalam +46.2%, Kannada +42.6% over Chandra 1).
  • Converts PDFs and images to structured Markdown, HTML, or JSON with full layout, bounding boxes, and image captions in a single pass.
  • Free for research and startups under $2M; available via GitHub and HuggingFace, with a hosted API offering $5 free credits.
  • Built on Qwen 3.5 VLM fine-tuned on diverse document types; processes ~2 pages/sec on an H100 in real-world usage.

Chandra OCR 2, from Datalab, is a 4-billion parameter open-source model that converts images and PDFs into structured Markdown, HTML, or JSON while preserving full layout information. It just claimed the top spot on the olmOCR benchmark, the most widely used OCR evaluation suite, scoring 85.9% overall and outperforming GPT-4o, Gemini Flash 2, Mistral OCR, and every other open-weight model in the field.

What makes this notable is not just the score. Chandra 2 is smaller than Chandra 1, but improves on it in every category. The model dropped from 9B to 4B parameters, while pushing the olmOCR benchmark score from 83.1% to 85.9% and multilingual accuracy from 69.4% to 77.8% across the top 43 languages. Fewer parameters, better results, and twice the throughput.

The problem it actually solves

Everyone is racing to build smarter AI agents, but there is a quiet problem upstream: most real-world data is trapped in documents. PDFs, scanned insurance forms, handwritten clinical notes, lab reports with complex tables , none of this flows cleanly into an LLM context window. Traditional OCR gives you a text dump. What agents actually need is structured output: tables that are still tables, form fields that are labeled, images that are described, layout that is preserved.

Chandra 2 is built specifically for that. It converts images and PDFs into structured HTML, Markdown, or JSON while keeping full layout information intact. It is not just extracting characters; it is reconstructing the document as a structured artifact.

One model, one pass

What distinguishes Chandra OCR from its predecessors is its full-page decoding approach. Traditional OCR systems operate as pipelines: first segment the document into blocks (text, table, image), then process each block separately. This approach works reasonably for simple documents but fails when layout becomes complex. A table whose merged cells span multiple columns, or a page mixing printed text and handwriting, overwhelms segmentation pipelines.