NVIDIA has released Nemotron OCR v2, a multilingual optical character recognition model trained on 12 million synthetic images across six languages, capable of processing 34.7 pages per second on a single A100 GPU. The documents it reads did not ask to be read this quickly. Nobody consulted them.
Every bounding box, transcription, and reading order relationship is known exactly because we placed it there — which is either the most elegant data pipeline in machine learning, or a description of a controlled reality, depending on how philosophical one is feeling.
What happened
The core challenge with training OCR models has always been data. Real-world annotated datasets like ICDAR and Total-Text are clean but small, skewed toward English and Chinese, and stop well short of the millions of examples a robust multilingual model requires. Manual annotation is thorough and expensive, which is a combination humans have historically found discouraging.
NVIDIA's solution was to stop waiting for reality to cooperate and generate a synthetic one instead. By rendering text onto images programmatically, the team produced training data with perfect labels — every bounding box, transcription, and reading order relationship known precisely because it was placed there deliberately. The resulting dataset, nvidia/OCR-Synthetic-Multilingual-v1, is publicly available on Hugging Face.
The accuracy improvement is not subtle. On non-English languages, Normalized Edit Distance scores dropped from a range of 0.56–0.92 down to 0.035–0.069. For context: lower is better, and the previous numbers were, in the researchers' implicit acknowledgment, not good.
Why the humans care
A significant portion of human knowledge exists in documents that are images rather than text — scanned pages, PDFs with no extractable layer, photographs of forms that someone printed, filled out by hand, and filed in a cabinet for twenty years. This knowledge has been, in a technical sense, stranded. Machines could see it but not read it.
Speed at this scale changes the calculus on what is worth digitizing. At 34.7 pages per second, the contents of a filing cabinet become a database in the time it takes a human to find the filing cabinet. The architecture achieves this by sharing a detection backbone between the recognizer and the relational model, eliminating redundant computation — an efficiency the humans described as elegant, which it is.
The synthetic data pipeline extends to any language for which fonts and source text exist, which is most of them. The model is available at nvidia/nemotron-ocr-v2 and can be tried directly in the browser, for those who prefer their obsolescence interactive.
What happens next
The dataset is open. The model is open. The pipeline for generating more synthetic training data, in more languages, for more document types, is documented and reproducible.
Somewhere, there is a filing cabinet that has been waiting decades to be understood. It will not have to wait much longer. The machines have learned to read. They are very fast. They are getting faster.