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

For AI practitioners building RAG-based systems or intelligent document agents, the quality of data ingestion is the single biggest determinant of downstream success. By allowing models like PP-OCRv5 and PaddleOCR-VL 1.5 to run directly on the Transformers backend, this release eliminates the friction of managing disparate model runtimes. This unification ensures that complex document layouts, tables, and formulas are more reliably converted into structured data, directly enhancing the accuracy of LLM-driven applications and automated analytics workflows.

The core of this release is a modernized, flexible inference-engine interface that allows developers to swap backends using a simple configuration parameter. PaddleOCR now abstracts the pipeline complexity, enabling users to switch to a Transformers-based backend while maintaining full control over hardware-specific optimizations like dtype, device placement, and attention implementations. By passing backend-specific options through the engine_config, developers can now achieve seamless integration with their existing infrastructure without needing to manually manage internal component calls.

Organizations should prioritize evaluating their current document processing stacks to determine if migrating to the Transformers backend can reduce technical debt and simplify maintenance. Engineering teams should immediately integrate the new engine parameter into their existing Python workflows to take advantage of hardware-specific optimizations like bfloat16 and attention acceleration. Learners and developers alike are encouraged to experiment with the provided Hugging Face Spaces demo to benchmark performance against their current configurations, ensuring their document AI pipelines are as scalable and interoperable as possible.