Alibaba Cloud Releases Open-Source Document Parsing Model OvisOCR2
Alibaba Cloud has released its OvisOCR2 document parsing model as open-source software. The model achieved a record score on the OmniDocBench benchmark, marking a significant advancement for end-to-end approaches.
Alibaba Cloud, a major player in China's tech industry, has launched OvisOCR2, an open-source model designed for document parsing. The model has set a new benchmark record on OmniDocBench v1.6, achieving a comprehensive score of 96.58. This marks the first time an end-to-end model has outperformed traditional pipeline methods in this domain.
Document parsing is crucial for converting unstructured data, such as PDFs and scans, into structured text, which is essential for applications like enterprise knowledge bases and RAG systems. Previously, the field was dominated by pipeline methods, which involve separate models for layout analysis and content recognition. This approach, while established, suffers from high maintenance costs and potential error accumulation.
OvisOCR2 employs an end-to-end architecture. It takes a document image as input and generates a structured Markdown representation in a single pass, covering text, formulas, tables, and visual elements in a natural reading order. This consolidated approach simplifies the process compared to multi-model pipelines.
The model is derived from the Qwen3.5-0.8B model and has undergone advanced training procedures. Alibaba Cloud has released OvisOCR2 under the Apache 2.0 license, ensuring compatibility with common inference frameworks like vLLM and seamless integration into the Qwen ecosystem.
Developers can access OvisOCR2 on platforms like Hugging Face and ModelScope. A technical report provides further details on the model's architecture and performance.