ATH-MaaS/OvisOCR2
OvisOCR2 by ATH-MaaS is a compact 0.8 billion parameter end-to-end vision-language model for page-level document parsing, post-trained from Qwen3.5-0.8B. It excels at generating Markdown representations of document pages, including text, formulas, tables, and visual regions, in natural reading order. The model achieves state-of-the-art performance on OmniDocBench v1.6 with a score of 96.58, becoming the first end-to-end model to lead this benchmark. OvisOCR2 is optimized for robust document parsing with a small deployment footprint, making it suitable for efficient and accurate document digitization.
Loading preview...
OvisOCR2: End-to-End Document Parsing Model
OvisOCR2, developed by ATH-MaaS, is a compact 0.8 billion parameter model designed for comprehensive page-level document parsing. It takes a document page image as input and generates a structured Markdown representation, preserving natural reading order and covering diverse content types such as text, formulas, tables, and visual regions.
Key Capabilities & Features
- End-to-End Document Parsing: Processes entire document pages to extract and structure content.
- Markdown Output: Generates a Markdown representation, including LaTeX for formulas and HTML for tables.
- State-of-the-Art Performance: Achieves an overall score of 96.58 on OmniDocBench v1.6, topping the leaderboard previously dominated by pipeline methods. It also scores 75.06 Avg3 on PureDocBench.
- Compact & Efficient: Built on Qwen3.5-0.8B, offering strong performance with a small deployment footprint.
- Advanced Training: Utilizes a multi-stage training recipe (SFT, RL, OPD) with a blend of real-world and synthetic data.
Good for
- Automated Document Digitization: Converting scanned documents or images into structured, editable formats.
- Information Extraction: Accurately extracting text, tables, and formulas from complex document layouts.
- Building Document Processing Pipelines: Serving as a core component for applications requiring robust document understanding and content generation.