Document OCR often breaks not on clean PDFs, but on the pages people actually need to process: skewed scans, curved sheets, photographs, and dense layouts. The key distinction here is treating recognition as document understanding—preserving structure and extracting fields, not just transcribing lines of text.
Key Capabilities
- Handles text, tables, formulas, charts, seals, and key information extraction, producing formats such as Markdown, HTML tables, LaTeX, and JSON. That makes it useful when downstream systems need more than searchable plain text.
- Supports Chinese and English documents, including scans, curved pages, and recaptured images. Its text-centric scope is more targeted than a general-purpose vision-language model.
- The model has 1.3B parameters. The complete parsing pipeline combines it with PP-DocLayoutV3 for layout detection; reported throughput is 1.2 pages per second on an RTX 5090. Results depend on hardware, backend, and document content.
- Reported benchmark scores include 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2. These provide useful reference points, but scores from different benchmarks measure different tasks and should not be compared directly.
Who It Suits
Great fit if you need locally deployable Chinese/English document parsing, structured extraction, or a compact OCR model that can run through common inference backends. Look elsewhere if you need broad visual reasoning, extensive language coverage beyond Chinese and English, or guaranteed performance independent of hardware and input quality; those are outside its stated focus.
