PaddleOCR 3.5: Running OCR and Document Parsing Tasks with a Transformers Backend

PaddleOCR 3.5 has been released with support for a Hugging Face Transformers inference backend, bringing OCR and document parsing capabilities closer to Hugging Face-centered workflows. According to the source, the update introduces a more flexible inference-engine interface where developers can select the backend using the engine parameter and configure backend-specific options through engine_config. While PaddleOCR continues to offer model series such as PP-OCRv5 for OCR and PaddleOCR-VL 1.5 for document parsing, the addition of the Transformers backend allows these models to fit naturally into environments already utilizing PyTorch and Transformers infrastructure for model loading, experimentation, and deployment. The release aims to address document ingestion challenges in RAG, Document AI, search, analytics, and agent applications by reducing integration friction when turning complex page layouts, PDFs, and scanned documents into structured data. For maximum throughput, the default paddle_static backend is typically still recommended. The integration was supported by Hugging Face engineers, including Anton Vlasjuk, Raushan Turganbay, and Yoni Gozlan, who assisted with pull requests, documentation, and reviews. Developers can test the new functionality via a live demo on Hugging Face Spaces.