OvisOCR3 tops OCRBench v2 at 0.8B via five-teacher distillation

OvisOCR3, a 0.8B-parameter model, scores 96.74 overall on OmniDocBench v1.6 and ranks first on OCRBench v2 with 65.9 on the English split and 61.3 on Chinese, according to a technical report posted to arXiv.

The recipe is distillation at scale. The authors describe "Multi-teacher on-policy distillation transfers expertise from five domain-specific 4B teachers into the shared model," followed by "mixed-task reinforcement learning with domain-specific rewards." The compact student is compared against both general-purpose vision-language models and specialist OCR systems.

The single checkpoint collapses five jobs: "document parsing, general OCR, key information extraction (KIE), document figure captioning, and cross-page table reconstruction in a single model," the paper reports. Beyond OmniDocBench and OCRBench v2, the report claims the highest overall scores on PureDocBench, Real5-OmniDocBench, and MDPBench, the highest localization F1 on DocFigCapBench, and the leading Relation F1 on CrossTabBench.

The abstract reports leaderboard numbers but does not name the five teacher models, does not publish inference latency versus the 4B teachers, and does not isolate how much of the gain comes from the RL stage versus the distillation step. The in-production test is whether a 0.8B unified model actually matches specialist systems on real documents, not only on curated benchmark splits.