PaddleOCR
An OCR and document-parsing toolkit that converts PDFs and images into structured Markdown or JSON, including a vision-language model for layout-aware document parsing and text recognition in more than 100 languages.
PaddleOCR converts PDF documents and images into structured data formatted as Markdown or JSON. Its PaddleOCR-VL vision-language model performs document parsing, including text, formulas, tables, and coordinates for table cells and text, and handles historical documents, rare characters, seals and charts; a separate PP-StructureV3 pipeline converts complex PDFs and images into Markdown or JSON with fine-grained coordinate information.
Its PP-OCRv6 scene-text model recognizes text in over 100 languages, including Chinese, English, Japanese and 46 Latin-script languages, in a single unified model, and also reads text in natural scenes such as street signs, IDs and industrial components. A related HPD-Parsing model offers high-throughput document parsing through either an OpenAI-compatible server or local inference with a customized vLLM runtime.
It integrates with RAG and agent projects such as Dify, RAGFlow, Pathway and Cherry Studio, supports NVIDIA GPUs, Intel CPUs, Kunlunxin XPUs and other AI accelerators, and is written in Python under the Apache-2.0 license.