NVIDIA Open-Sources Nemotron OCR v2 — Multilingual OCR at 34.7 Pages/Sec on A100, Trained on Synthetic Data

Hugging Face Blog·medium signal

NVIDIA released Nemotron OCR v2 on Hugging Face under the NVIDIA Open Model License — a production-ready multilingual OCR system supporting English, Chinese (Simplified/Traditional), Japanese, Korean, and Russian. The hybrid architecture (detector + recognizer + relational model) processes 34.7 pages/second on a single A100 by sharing a detection backbone across all three modules. Trained primarily on synthetic data plus ~680K real-world images. For RAG pipeline builders: this is a serious free alternative to commercial OCR services.