NVIDIA Releases Nemotron-Parse-2.0 for Advanced Document Parsing and Layout Analysis
NVIDIA has released Nemotron-Parse-2.0, an upgraded vision-language model designed to convert document images into structured text and layout annotations. This version introduces chart-aware parsing, improved handwritten-text extraction, and a 20k-token vocabulary expansion for enhanced multilingual support. The model significantly improves document processing and Retrieval-Augmented Generation (RAG) pipelines by accurately parsing complex, chart-heavy, and multilingual documents. Its commercially permissive license allows developers to freely integrate it into enterprise AI agents and data curation workflows. Nemotron-Parse-2.0 utilizes a new `<class_Chart>` token to identify chart regions and convert visual chart data into structured text. It also offers enhanced OCR performance for CJK (Chinese, Japanese, Korean) and Indic scripts, alongside stronger table structure recovery.
## BACKGROUND
Document layout analysis is a computer vision and NLP process that identifies and categorizes regions of interest, such as text, tables, and images, in scanned documents. Previous versions like Nemotron-Parse v1.2 focused on basic semantic understanding and spatial grounding, but struggled with complex charts and diverse multilingual scripts.