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

PaddleOCR 3.5 introduces a flexible inference-engine interface, allowing its Optical Character Recognition (OCR) and document parsing models to operate with Hugging Face Transformers as an inference backend. This integration provides developers with more choices for running these capabilities while retaining PaddleOCR’s established model series. The update is particularly beneficial for those already utilizing a Hugging Face-centered stack.

Before large language models (LLMs) can effectively process information, raw data from various document formats—such as PDFs, scanned images, or complex layouts—must first be converted into reliable structured data. This initial ingestion step is crucial, as weaknesses here can lead to LLMs missing key information or generating inaccurate responses.

PaddleOCR addresses this challenge by offering robust OCR models like PP-OCRv5 and document parsing models such as PaddleOCR-VL-1.5. By enabling these models to connect with Transformers-centered stacks, PaddleOCR 3.5 simplifies the overall workflow from document ingestion to downstream applications like RAG, search, and analytics.

Developers can now seamlessly integrate PaddleOCR’s powerful capabilities within their existing Transformer-based infrastructures. While PaddleOCR continues to manage the core OCR and document parsing pipelines, the option to use Transformers as a backend offers reduced integration friction and a more natural development path for advanced Document AI applications. However, for maximum OCR or document parsing throughput, PaddleOCR’s default paddle_static backend generally remains the recommended choice, highlighting that this release is about enhancing flexibility rather than replacing existing backends.