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

PaddleOCR Integrates Transformers for Document AI Workflows

PaddlePaddle has released PaddleOCR 3.5, introducing support for Hugging Face Transformers as an optional inference backend. This update is aimed at developers working within the PyTorch ecosystem, as it simplifies the process of integrating sophisticated OCR and document parsing capabilities into applications like Retrieval-Augmented Generation (RAG) and Document AI. By allowing its models to run on a Transformers runtime, PaddleOCR lowers a key adoption barrier for teams that need to process complex documents, such as PDFs and scanned images, before feeding structured data to a large language model.

Technical Details and Developer Control

The integration is enabled through a new, more flexible inference interface. Developers can now switch to the new backend by setting a single parameter, `engine="transformers"`. While PaddleOCR continues to manage the underlying document processing pipelines, this release gives users more control over the execution environment. Key technical features include:

• Model series like PP-OCRv5 and PaddleOCR-VL 1.5 are now compatible with the Transformers backend.
• A new `engine_config` parameter allows for backend-specific tuning, including `dtype`, device placement, and attention implementation (`sdpa`).
• The default `paddle_static` backend remains available and is recommended for scenarios where maximizing throughput is the primary objective.

This layered approach allows application builders to leverage PaddleOCR's models without needing to manually call each internal component, while still providing the configuration options needed for custom deployments. The change positions Transformers as an alternative runtime, not a replacement, giving developers the ability to choose the backend that best fits their existing infrastructure.

Ecosystem Impact and Interoperability

By bridging its technology with the Hugging Face ecosystem, PaddlePaddle is making its specialized tools more accessible to a broader audience. This move addresses a common friction point in building complex AI systems, where data ingestion from unstructured documents is often a critical but difficult first step. For developers already using Transformers for model management and deployment, this integration creates a more natural path from raw documents to downstream analytics, search, and automation workflows. It reflects a strategic focus on interoperability, prioritizing developer experience and component flexibility over forcing users into a single, monolithic framework.

PaddlePaddle's integration of a Transformers backend for PaddleOCR is a pragmatic move to increase adoption by meeting developers where they are. Instead of competing for framework dominance, it positions its specialized OCR technology as a modular, high-value component within the larger, PyTorch-centric AI ecosystem.