OvisOCR2: A 0.8B End-to-End Document Parser Takes the OmniDocBench Crown From Pipeline Approaches (96.58)

arXiv / HuggingFace Daily Papers·medium signal

A compact 0.8B-parameter parser that emits Markdown covering text, formulas, tables, and visual elements, scoring a state-of-the-art 96.58 overall on OmniDocBench v1.6 — notable because it puts an end-to-end model atop a leaderboard historically led by multi-stage pipelines — plus a top Avg3 of 75.06 on PureDocBench. Open weights on HuggingFace (ATH-MaaS/OvisOCR2) under CC BY 4.0. For builders, a 0.8B model beating pipelines means local, cheap document ingestion is now credible.