Datalab has released Surya OCR 2, a new multilingual OCR model achieving 83.3% accuracy on the olmOCR-bench, setting a new standard for models under 3 billion parameters. Despite having only 650 million parameters, Surya OCR 2 outperforms its predecessor, which had 9 billion parameters, by achieving a Pareto optimal balance between parameter count and accuracy. The model integrates layout analysis, text recognition, and table detection into a single vision-language model, while maintaining separate lightweight models for text line detection and OCR error detection. Surya OCR 2 supports 91 languages with an overall pass rate of 87.2% and features optimizations for damaged documents and handwritten text. It offers high deployment efficiency, achieving 5.35 pages per second on NVIDIA GPUs and supporting local inference on Apple M1 devices. The model is open-sourced under the Apache 2.0 license, with weights available under the OpenRAIL-M license. Datalab also introduced a paid API for the enhanced 4-billion-parameter Chandra 2 model.