PaddleOCR is an OCR toolkit and document AI engine from the PaddlePaddle project that converts PDFs and images into structured Markdown or JSON for use in retrieval-augmented generation and agent pipelines. It covers scene text recognition, layout-aware document parsing, and recognition of tables, formulas, seals, and charts.
Its model families serve different jobs. PP-OCR handles text detection and recognition: PP-OCRv6 recognizes 50 languages (Chinese, English, Japanese, and 46 Latin-script languages) with one unified model, in tiny (1.5M), small (7.7M), and medium (34.5M) parameter tiers for edge, mobile, and server use. PP-StructureV3 converts complex documents to Markdown or JSON with fine-grained coordinates for table cells and text. The PaddleOCR-VL series is a 0.9B vision-language model for page-level document parsing across more than 100 languages, and HPD-Parsing is a lightweight model built for high-throughput parsing.
Inference runs on Paddle static or dynamic graphs or on Transformers, can be accelerated with OpenVINO, ONNX Runtime, or TensorRT, and scales across multiple GPUs and processes. Models export to ONNX, a serving layer lets C++, C#, Java, and other languages call it over HTTP, a browser SDK runs PP-OCRv5 client-side, and hardware backends include NVIDIA GPUs, Intel CPUs, and Kunlunxin XPUs. A hosted website offers an interactive experience center and APIs.
| Model | Reported result |
|---|---|
| PaddleOCR-VL-1.6 (0.9B) | 96.3% on OmniDocBench v1.6 |
| PaddleOCR-VL-1.5 | 94.5% on OmniDocBench |
| PP-OCRv6 medium vs PP-OCRv5 server | +4.6% detection, +5.1% recognition |
| PP-OCRv6 inference | 5.2x CPU speedup with OpenVINO, 0.13s on an A100 GPU |
| HPD-Parsing | Peak throughput of 4,752 tokens/s |
Features
- Document parsing: PDFs and images converted to Markdown or JSON with layout, tables, formulas, seals, and charts
- Scene text recognition: IDs, street views, books, industrial parts, digital displays, and dot-matrix text
- Multilingual OCR: 50 languages in one PP-OCRv6 model and more than 100 across the model families
- Coordinate output: table cell, text, and single-character coordinates
- Office conversion: Word, Excel, and PowerPoint files to Markdown, and parsed results exported to DOCX
- Long documents: cross-page table merging and hierarchical heading detection
- Inference backends: Paddle, Transformers, OpenVINO, ONNX Runtime, TensorRT, and a customized vLLM runtime
- Model tiers: tiny, small, and medium models for edge, mobile, and server deployment
- Serving: HTTP serving for C++, C#, Java, and other languages, plus a browser SDK
- Chart to table: chart images converted into tables
- Benchmarking: per-layer and per-module latency measurement for each pipeline
