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

Hugging Face Blog · 1 month ago

PaddleOCR 3.5 now allows its OCR and document parsing models to run using Hugging Face Transformers as an inference backend by setting the engine parameter to "transformers". The release introduces a flexible inference-engine interface where developers can select backends and configure backend-specific options like dtype, device placement, and attention implementation through an engine_config parameter. This integration reduces friction for developers building RAG, Document AI, and agent applications that already rely on PyTorch and Transformers infrastructure, while the paddle_static backend remains the recommended choice when maximizing throughput is the priority.