TurboOCR: 270–1200 img/s OCR with Paddle and TensorRT (C++/CUDA, FP16)

Docling and MinerU are great for structured output like markdown and table extraction, but they run at 1-5 pages/s because of the VLMs under the hood.

Turbo-OCR gives you bounding boxes, text, and layout regions at multiple hundred img/s depending on the text density. When you have many PDFs to process, it makes a huge difference. You can always pipe the output into a VLM for the pages that need deeper extraction. Structured extraction and markdown output are on the roadmap (without sacrificing too much speed).

Thanks, happy to hear that!