650 million parameters run to 83.3%, Surya OCR sets a new benchmark for open source OCR below 2-3B According to Beating monitoring, the open-source document intelligence platform Datalab has officially released the new multilingual OCR open-source model Surya OCR 2. The new model has only 650 million parameters and achieved 83.3% in the authoritative document intelligent evaluation olmOCR bench, ranking first in the level below 3 billion parameters. Its performance even surpasses the first generation 9 billion parameter version with a volume of about 14 times, achieving Pareto optimality in parameter quantity and accuracy. In terms of functionality, Surya OCR 2 consolidates the three major tasks of layout analysis, text recognition, and table recognition into a single visual language model (VLM), while text line detection and OCR error detection still run through independent lightweight models. Users can complete full page OCR recognition with a single model call, outputting structured HTML code containing coordinate boxes and reading order. Mathematical formulas are output in HTML math tags, while cross row and cross column tables are organized into standard HTML format. In terms of multilingual support, the new model achieved a comprehensive pass rate of 87.2% in 91 language tests (Chinese pass rate of 82.5%), and deeply optimized damaged documents and handwriting. In terms of deployment efficiency, Surya OCR 2 supports two major inference backends. The system runs Docker on Nvidia GPU devices with vLLM backend enabled, and a single RTX 5090 graphics card can achieve a super high throughput of 5.35 pages per second. In Apple devices or regular CPU environments, the system loads GGUF format through llama.cpp and achieves full local side operation on M1 computers. At present, the source code of the new model is open sourced based on the Apache 2.0 protocol, and the weight is provided free of charge to individuals, universities, and startups with an annual income of less than $5 million based on the OpenRAIL-M protocol. And Datalab has also officially opened a paid API with a stronger 4 billion parameter Chandra 2 model, and given away a $5 experience credit. [Original link]