Aliyun Open Sources 0.8B Document Parsing Model OvisOCR2, Ends-to-End Solution Tops OmniDocBench

On July 24, Alibaba announced the open-source release of OvisOCR2, a document parsing model with a parameter scale of only 0.8B. In the authoritative document parsing benchmark OmniDocBench v1.6 evaluation, this model achieved a comprehensive score of 96.58, ranking first and becoming the first end-to-end document parsing model to comprehensively surpass traditional pipeline methods, marking a breakthrough paradigm shift in the field of document intelligence technology.

As a core infrastructure for RAG retrieval, intelligent Q&A, and enterprise knowledge bases, document parsing has long been dominated by the "layout analysis + content recognition" pipeline structure, facing bottlenecks such as high maintenance costs, cumulative errors, and complex deployment. OvisOCR2, based on Qwen3.5-0.8B, overcomes these limitations, enabling single-model generation to output a Markdown representation containing text, formulas, tables, and visual areas in natural reading order.

In terms of technical implementation, the model combines complementary approaches using real and synthetic data, and fully unleashes the potential of compact models through four key methods: supervised fine-tuning (SFT), reinforcement learning (RL), online policy distillation (OPD), and model ensembling.