NVIDIA Unveils Document-Specialized VLM

Llama Nemotron Nano VL is an 8B-scale Vision Language Model (VLM) optimized for Intelligent Document Processing (IDP) and OCR.

It specializes in accurately extracting and understanding text, tables, charts, and diagrams from complex documents such as invoices, receipts, and contracts. It has demonstrated high performance on the OCRBench v2 benchmark, and offers the following key features:

• Text Recognition and Element Parsing: Accurately identifies text and key elements such as tables, charts, and images within documents.
• Table Extraction: Extracts complex tabular data, such as financial statements, with high accuracy.
• Grounding: Supports bounding boxes for both queries and outputs, enhancing the interpretability of model responses.

This model is built on Llama-3.1-8B-Instruct and the C-RADIOv2-VLM-H Vision Transformer for high-resolution visual feature extraction. Users can use the model immediately on Hugging Face, and can also perform additional post-training on their own datasets using NVIDIA NeMo.