TurboOCR v3 adds structured parsing and reaches 520 img/s on RTX 5090

TurboOCR v3 is a self-hosted, high-speed document OCR server that now supports end-to-end structured parsing while maintaining full local execution via HTTP and gRPC.

• The pipeline switches to PP-OCRv6 models, increasing throughput from ~270 to ~520 img/s on FUNSD using a tiny model on an RTX 5090.
• New structured parsing converts layout to HTML tables, formulas to LaTeX, and reading order to Markdown.
• Table and formula processing is strictly opt-in per request to avoid unnecessary computational costs.
• The stack utilizes C++, TensorRT FP16, multi-streaming, and includes a direct PDF endpoint.