NVIDIA Releases Nemotron OCR v2 Multilingual Model Trained on 12 Million Synthetic Images

NVIDIA has introduced Nemotron OCR v2, an advanced multilingual optical character recognition model capable of processing 34.7 pages per second using a single A100 GPU. This model accommodates six languages: English, Japanese, Korean, Russian, Chinese, and an undisclosed language. It was trained on a dataset comprising 12 million synthetic images derived from the mOSCAR multilingual web corpus. Featuring a shared detection backbone (RegNetX-8GF), it enhances Normalized Edit Distance scores for non-English languages from 0.56–0.92 to 0.035–0.069. Character support has increased from 855 to 14,244, now including CJK and Cyrillic scripts. The dataset can be found at nvidia/OCR-Synthetic-Multilingual-v1, while the model is available at nvidia/nemotron-ocr-v2. Key contributors include Bo Liu, Théo Viel, and Mike Ranzinger.