Eagle is a research project developing vision-language models that combine visual understanding with large language model capabilities. The project includes Eagle, Eagle 2, and Eagle 2.5 versions, with models available on HuggingFace. These models serve as the VLM backbone for NVIDIA’s robotics foundation models like GR00T, and Eagle 2.5 was accepted to NeurIPS 2025. The models are built on Llama architectures with data-centric training strategies focused on data quality and curation.