The 0.9B vision model that actually reads your PDFs

What it does PaddleOCR is an open-source toolkit that extracts text, tables, formulas, and layout structure from images and PDFs, converting the mess into structured Markdown or JSON for downstream AI pipelines. It covers everything from street-view text recognition to parsing complex documents with seals and rare scripts. The project also handles layout analysis, document translation, and runs on hardware from consumer CPUs to dedicated AI accelerators.

The interesting bit Rather than chaining a dozen specialist models, the project fields a single 0.9-billion-parameter vision-language model—PaddleOCR-VL-1.6—that claims 96.3% accuracy on OmniDocBench v1.6 and supposedly outperforms larger proprietary solutions. That is an unusually small footprint for a parser that also handles charts, ancient documents, and seal recognition.

Key highlights

• PaddleOCR-VL-1.6 (0.9B params) parses page-level documents into Markdown or JSON, including tables, formulas, seals, and charts.
• PP-OCRv5 supports 100+ languages in a single model, handling mixed multilingual documents.
• PP-StructureV3 extracts fine-grained layout coordinates—table cells, text blocks—for downstream processing.
• Integrations with Dify, RAGFlow, and Cherry Studio; parsed results export to DOCX; and PaddleOCR.js runs PP-OCRv5 directly in the browser.
• Hardware backends include NVIDIA GPU, Intel CPU, Kunlunxin XPU, and various NPUs.

Verdict Worth evaluating if you are building RAG or agentic pipelines that need structured data from messy scans. Probably overkill if you just need to extract clean text from a handful of screenshots.