NVIDIA has launched the Llama Nemotron Nano VL, an 8‑billion‑parameter vision‑language model now hosted on the Hugging Face Hub. Designed for intelligent document processing, the model combines Llama‑3.1‑8B‑Instruct with a Vision Transformer backbone, C‑RADIOv2‑VLM‑H, to deliver high‑accuracy OCR, table extraction, and visual reasoning on complex documents such as invoices, receipts, contracts, and academic papers. Training leveraged a diverse blend of synthetic and real datasets, totaling roughly 1.5 million samples, and included public collections like DocLayNet, FinTabNet, and PubTables‑1M. The model achieved industry‑leading performance on the OCRBench v2 benchmark, outperforming current leading VLMs in real‑world enterprise scenarios. NVIDIA emphasizes the model’s low footprint and efficient deployment, making it suitable for fast, scalable document‑workflow automation across finance, healthcare, legal, and government domains. Users can fine‑tune the model on their own data using NVIDIA NeMo, and tutorials are available to help integrate the Llama Nemotron Nano VL into existing pipelines.
Welcome the NVIDIA Llama Nemotron Nano VLM to Hugging Face Hub
NVIDIA has released the Llama Nemotron Nano VL, an 8B vision‑language model now available on Hugging Face Hub. The model is tailored for intelligent document processing, excelling in OCR, table extraction, and visual reasoning on complex documents such as invoices and contracts. It builds on Llama‑3.1‑8B‑Instruct and a Vision Transformer backbone, C‑RADIOv2‑VLM‑H, and was trained on a mix of synthetic and real datasets totaling about 1.5 million samples. The model achieves high accuracy on the OCRBench v2 benchmark and supports fast, scalable deployments.
Why it matters
The release demonstrates NVIDIA’s continued push into multimodal AI, offering a compact, high‑accuracy model that can be deployed at scale for enterprise document workflows, potentially reducing costs and improving automation across finance, legal, and government sectors.