AI BriefingPaddleOCR 3.5 Adds Transformers Backend Support
Key point
PaddleOCR 3.5 has added Transformers backend support.
Details
PaddleOCR 3.5 now lets you run supported models on the Hugging Face Transformers backend via engine=transformers.
The inference engine interface has become more flexible: you pick the backend with engine, then tune dtype, device placement, and attention implementation with engine_config. Since PaddleOCR manages the internal pipeline, users don't need to call each component separately.
- The model layer keeps PP-OCRv5 and PaddleOCR-VL 1.5.
- The backend layer includes Paddle static graph, Paddle dynamic graph, and Transformers.
- It makes it easier to attach the document input stage in RAG, Document AI, search, and agent workflows.
The installation example, based on CUDA 12.6, guides you to match paddleocr==3.5.0, paddlex==3.5.2, transformers>=5.4.0 with a compatible PyTorch build. In the CLI, run with --engine transformers, and the same configuration can be passed via the Python API as well.
If throughput is the top priority, the default paddle_static backend is still recommended, and a Hugging Face Spaces demo and Hub model pages have also been released together.
This summary was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report summary errors, attribution issues, or removal requests via Contact.