Hugging Face AI Updates: April 18, 2026

1. NVIDIA Drops Nemotron OCR v2: 34.7 Pages/Sec on a Single A100 via Synthetic Training

Hugging Face. NVIDIA published Nemotron OCR v2 on the Hub, a multilingual OCR model that hits 34.7 pages per second on a single A100 while covering English, Chinese Simplified and Traditional, Japanese, Korean, and Russian in one model. The unified three-component architecture — RegNetX-8GF detector, Transformer recognizer, and relational model sharing a convolutional backbone — was trained on 12.2M synthetic images generated from the mOSCAR corpus and open-source fonts, achieving Normalized Edit Distance scores of 0.035-0.069 on non-English languages. On OmniDocBench the multilingual variant runs 28x faster than PaddleOCR v5 at competitive accuracy. Source