🔗 RIP commercial OCR. An open-source model just topped every benchmark.

Chandra OCR 2, a model from Brooklyn's Datalab, recently made waves in the OCR realm. With just 4 billion parameters, it outperformed both GPT-4o and Gemini across 90 languages, showcasing a remarkable ability to decode complex layouts and produce structured formats like Markdown and HTML. In independent benchmarks, it scored a staggering 85.9%, highlighting a clear edge over well-known models, diverging from typical OCR tools that struggle with intricacies like handwritten notes or multi-column documents. The article by Sumit Pandey emphasizes the implications of Chandra's performance, suggesting it could reframe expectations in document processing pipelines and beyond. The model’s open-source nature and easy accessibility, paired with its remarkable accuracy, means it’s poised to be a game-changer for anyone working with tricky, real-world documents.