TurboOCR v3 makes local document parsing faster and more structured
TurboOCR v3 is a tool that reads text and structure from document images. The new version moves to PP-OCRv6 models and reports a speed increase on the FUNSD document dataset from about 270 images per second to about 520 images per second. That is roughly 1.9 times faster on the same RTX 5090 setup with the same dataset and measurement, while keeping similar accuracy.
It still runs fully local and does not use a VLM. The bigger feature change is structured from start to finish. It can turn tables into HTML, formulas into LaTeX, and the document’s reading order into Markdown.
Table and formula parsing are only turned on when requested. The main limits are hardware and startup time: it is NVIDIA-only because it uses TensorRT, and the first startup can take hours while the runtime engines are built, though later starts are fast because the engines are cached. It serves HTTP and gRPC from one .
Key points
- PP-OCRv6 raises reported from about 270 to about 520 images per second.
- The speed gain is about 1.9 times while keeping similar accuracy.
- It and does not rely on a VLM.
- It can output tables as HTML, formulas as LaTeX, and reading-order text as Markdown.
- It requires NVIDIA hardware through TensorRT, and the first startup can take hours.