PaddleOCR 3.5 now runs on a Transformers backend, cutting a major friction point for RAG devs

The team behind PaddleOCR just shipped version 3.5, and the real news isn’t a new model—it’s a new inference engine option. You can now run PaddleOCR’s document parsing and OCR pipelines, including PP-OCRv5 and PaddleOCR-VL 1.5, directly on a Hugging Face Transformers backend by simply setting engine="transformers".

This is a big deal for developers who’ve been stitching together document-heavy RAG applications. The hard part of document AI almost always comes before the LLM even touches the data. You’re dealing with PDFs, scanned forms, screenshots, tables, and formulas, and if the ingestion step is sloppy, the downstream agent or search workflow falls apart. PaddleOCR already solved a lot of that extraction pain, but it often meant working outside the PyTorch/Transformers ecosystem that many teams have standardized on for model loading and deployment.

Now that friction is gone. You can install it alongside your existing stack (pip install paddleocr==3.5.0 transformers>=5.4.0), pass engine-specific configs like dtype and attention implementation through engine_config, and keep everything in one familiar environment. The Hugging Face demo space uses float32 for broad compatibility, but you can tune it on your own hardware with bfloat16 or sdpa attention where it makes sense.

Let’s be clear about what this isn’t. It’s not a performance play to replace PaddlePaddle’s native paddle_static backend, which remains the recommended choice when throughput is everything. This is a flexibility play. If you’re already all-in on PyTorch infrastructure and just want solid OCR without the integration headache, PaddleOCR 3.5 now meets you where you are. The team gave a specific shout-out to Hugging Face engineer Anton Vlasjuk for shepherding the pull requests through, which tells you this was a real collaboration, not a throwaway integration.