End-to-End Model Surpasses Pipeline for First Time: What Does OvisOCR2's 96.58 Score Mean?

96.58 points—that's the comprehensive score Alibaba Cloud's OvisOCR2 achieved on the OmniDocBench v1.6 leaderboard. 0.8B parameters—that's the size of the model. Before it, the top ten on the leaderboard were all pipeline methods—meaning detection, recognition, layout analysis, and structured output were split into multiple independent modules processed separately. OvisOCR2 is the first model to reach the top spot with an end-to-end architecture, and its parameter scale is under 1B.