Freedium

Open original

Chandra OCR 2, a 4 billion-parameter model from Datalab, outperforms GPT-4o and Gemini on AllenAI’s olmOCR benchmark and a 90-language test while halving the model size. It preserves layout, reads complex tables and math notation, converts diagrams to Mermaid, and runs at two pages per second on an NVIDIA H100. The code is Apache 2.0 but the model weights use an OpenRAIL-M license with commercial restrictions.

  • Chandra OCR 2 scores 85.9% on olmOCR vs GPT-4o's 69.9%, while cutting model size from 9B to 4B parameters and hitting ~2 pages/sec on an H100
  • Multilingual performance beats Gemini 2.5 Flash and GPT-5 Mini, with 40-46 point gains on scripts like Kannada, Malayalam and Telugu over Chandra 1
  • Weights carry an OpenRAIL-M license requiring a paid commercial license once a company exceeds $2M in funding or revenue, despite Apache 2.0 code
  • Handwriting recognition remains weak, dropping to ~50.4% accuracy on complex forms despite strong printed-text and table/math handling

Datalab’s new Chandra OCR 2 is a four-billion-parameter open-weight model that turns images and PDFs into structured Markdown, HTML or JSON while keeping the original layout intact. On the independent olmOCR benchmark it scored 85.9%—16 points higher than GPT-4o’s 69.9%. In Datalab’s own 90-language test, Chandra hit 72.7% versus Gemini 2.5 Flash at 60.8%, and on the 43 most common languages it jumped to 77.8%, outpacing GPT-5 Mini’s 60.5%. Scripts like Kannada, Malayalam and Telugu saw gains of 40–46 points compared to Chandra 1. Beyond printed English, Chandra OCR 2 handles complex tables with nested headers, handwritten math on century-old scans, multi-column layouts, forms with checkboxes, images with auto-captions and even exports flowcharts as Mermaid diagrams. They halved model size from 9B to 4B parameters, boosted accuracy across all categories and doubled throughput—about two pages per second on a single NVIDIA H100. Installation is a five-line pip command with no API keys or per-page fees, or you can deploy via Docker and vLLM for production. Licensing is Apache 2.0 for code, but weights use a modified OpenRAIL-M license: free for research, personal use and startups under $2 million in funding or revenue; larger companies need a commercial license. The olmOCR results come from an independent benchmark, but Datalab’s multilingual scores await third-party validation. GPT-4o’s score also reflects fixed prompts, so fine-tuned prompting could narrow the gap. Handwriting only hits around 90.8% on notes and falls to about 50.4% on complex forms, so heavy handwriting use still needs testing.

Ask about this article