Eagle
A family of frontier vision-language models from NVIDIA exploring data-centric training strategies for general multimodal understanding, long-context reasoning and embodied applications.
Eagle is a family of vision-language models (VLMs) from NVIDIA that explores data-centric training strategies across general-purpose multimodal understanding, long-context reasoning and embodied, physical-world applications. Beyond its own research results, Eagle has served as the underlying research and development platform for several other NVIDIA efforts, including Llama-Nemotron Nano VLM, Nemotron VLMs, NeMo Retriever's Llama Nemoretriever Colembed model, and the Isaac GR00T N1 family of embodied-AI models.
The family includes LocateAnything, a generalist vision-language grounding, detection and pointing model that supports batch inference through a FlashAttention runtime and can be fine-tuned with LoRA, and Eagle 2.5, a frontier VLM aimed at long-context image and video understanding; Eagle 2.5's methods were accepted at NeurIPS 2025, and LocateAnything's at ECCV 2026. A native-resolution variant of the model is also used as the VLM backbone of GR00T-N1.6.
It is written in Python and licensed under Apache-2.0.