Eagle
Dataset & ModelNVlabs/Eagle
Frontier vision-language models with data-centric strategies.
Overview
Eagle is a family of frontier vision-language models from NVIDIA, exploring data-centric strategies for multimodal understanding, long-context reasoning, and embodied AI. It includes multiple versions (Eagle, Eagle 2, Eagle 2.5) and supports tasks like generalist grounding, detection, and video understanding.
README Preview
\n\n# 🦅 Eagle: Frontier Vision-Language Models with Data-Centric Strategies\n\n\n \n\n\n[](LICENSE)\n[](./Eagle2_5/LICENSE_MODEL)\n\n[[📘Eagle Report](Eagle/Eagle.pdf)] [[📘Eagle 2 Report](Eagle2_5/Eagle2.pdf)] [[📘Eagle 2.5 Report](Eagle2_5/Eagle2.5.pdf)] [[📘LocateAnything Report](https://research.nvidia.com/labs/lpr/locate-anything/LocateAnything.pdf)]\n\n[[🤗Model Collection](https://huggingface.co/collections/nvidia/eagle)] [[🤗LocateAnything Demo](https://huggingface.co/spaces/nvidia/LocateAnything)] [[🌐Project Page](https://nvlabs.github.io/Eagle/)]\n\n\n\n\n## Updates\n- [2026/06] 🎉 LocateAnything is accepted to [ECCV 2026](https://eccv.ecva.net/).\n- [2026/06] 🔥 Release [visual prompt fine-tuning script](./Embodied/shell/locate-anything-lora-visual-prompt.sh) for LocateAnything with LoRA fine-tuning.\n- [2026/06] 🔥 LocateAnything now supports [batch inference](./Embodied/) with a pure FlashAttention runtime — efficient inference on A100, RTX 4090, and other non-Hopper/Blackwell GPUs.\n- [2026/05] 🔥 Release [LocateAnything](./Embodied/) — A generalist vision-language grounding model based on Eagle.\n- [2025/12] 🎉 A native resolution variant of the Eagle model is adopted as the VLM backbone of [GR00T-N1.6](https://huggingface.co/collections/nvidia/gr00t-n16). Check out the [tech blog](https://research.nvidia.com/labs/gear/gr00t-n1_6/) for more details.\n- [2025/10] 🔥 Release Eagle 2.5 [source code](https://github.com/NVlabs/EAGLE/tree/main/Eagle2_5).\n- [2025/09] 🔥 Eagle 2.5 is accepted to [NeurIPS 2025](https://neurips.cc/Conferences/2025).\n- [2025/09] 🎉 Eagle 2 is supported in [Torch-TRT](https://github.com/pytorch/TensorRT/tree/main/tools/llm).\n- [2025/07] 🎉 Release Eagle 2.5 [model](https://huggingface.co/nvidia/Eagle2.5-8B).\n- [2025/06] 🔥 Eagle 2.5 is adopted as the VLM backbone of [GR00T-N1.5](https://huggingface.co/nvidia/GR00T-N1.5-3B). Check out the [tech blog](https://research.nvidia.com/labs/gear/gr00t-n1_5/) for more details.\n
FAQ (1)
TroubleshootingHow to fix Transformers ImportError: “This modeling file requires the following packages that were not found in your environment: peft” even after running pip install peft?
The package is likely installed for a different Python interpreter, the install is incomplete, or the latest PEFT is incompatible with Transformers 4.37.2. In the activated eagle conda environment, install PEFT using the same interpreter that runs inference:
which python
python -m pip --version
python -m pip install --force-reinstall --no-cache-dir peft
python -c "import peft; print(peft.__version__, peft.__file__)"If the import fails or the error remains with PEFT 0.15.x, install a release contemporary with Transformers 4.37.2:
python -m pip install "peft==0.9.0"Then restart the Python process and rerun inference.py. Verify both packages are visible to the same interpreter with:
python -m pip list | grep -E "peft|transformers"