TurboOCR

TurboOCR is a self hosted document parsing server built on NVIDIA TensorRT. It runs PaddleOCR PPOCRv6 weights in a native C++ runtime and packages everything in a single container that serves HTTP and gRPC. Beyond plain text recognition it performs layout analysis, converts tables to HTML and formulas to LaTeX, handles PDFs natively and exports results as Markdown. It is built for engineering teams that need high throughput OCR without sending documents outside their network. The vendor reports up to 556 images per second for whole page OCR on receipts using one RTX 5090 GPU, about 20 pages per second for full parsing with layout, tables and formulas, and a 92 percent word F1 score on the FUNSD benchmark. A request such as POST /ocr/raw returns JSON with each text line, a confidence value and a four point bounding polygon. Version 3.2 added layout aware OCR, native PDF support and TensorRT acceleration. A hosted demo endpoint at api.turboocr.com lets visitors try the API, and the documentation is available in English, Chinese, German and French. The project is backed by MiruIQ. TurboOCR is open source under the MIT license and the code is on GitHub. It requires an NVIDIA GPU with CUDA. The site publishes no pricing and lists no paid plans.