Krux
April 19, 2026
Nvidia's New OCR Reads 35 Pages Per Second in Five Languages
Published: April 19, 2026 at 12:38 AM
Updated: April 19, 2026 at 12:38 AM
100-word summary
Nvidia released Nemotron OCR v2, a single model that reads English, Chinese, Japanese, Korean, and Russian documents without needing separate language-specific models. The upgrade jumps from 855 characters to 14,244, handling everything from alphabets to ideographs in one pass. The speed hook: 34.7 pages per second on a single A100 GPU. That makes bulk digitization of mixed-language archives suddenly feasible for organizations that used to juggle multiple OCR tools. The twist? Nvidia trained it almost entirely on synthetic data, generating millions of annotated pages per day from fonts and text corpora. Adding new languages now requires source text and typefaces, not hand-labeled scans.
What happened
Nvidia released Nemotron OCR v2, a single model that reads English, Chinese, Japanese, Korean, and Russian documents without needing separate language-specific models. The upgrade jumps from 855 characters to 14,244, handling everything from alphabets to ideographs in one pass. The speed hook: 34.7 pages per second on a single A100 GPU. That makes bulk digitization of mixed-language archives suddenly feasible for organizations that used to juggle multiple OCR tools.
Why it matters
The twist? Nvidia trained it almost entirely on synthetic data, generating millions of annotated pages per day from fonts and text corpora. Adding new languages now requires source text and typefaces, not hand-labeled scans.