Eagle

🦅 Eagle: Frontier Vision-Language Models with Data-Centric Strategies

[📘Eagle Report] [📘Eagle 2 Report] [📘Eagle 2.5 Report] [📘LocateAnything Report] [📘VFlash Report]

[🤗Model Collection] [🤗LocateAnything Demo] [🌐Project Page]

Updates

• [2026/11] 🔥 Release VFlash: fast, accurate visual planning via Flash Primitive Decoding.
• [2026/06] 🎉 LocateAnything is accepted to ECCV 2026.
• [2026/06] 🔥 Release visual prompt fine-tuning script for LocateAnything with LoRA fine-tuning.
• [2026/06] 🔥 LocateAnything now supports batch inference with a pure FlashAttention runtime — efficient inference on A100, RTX 4090, and other non-Hopper/Blackwell GPUs.
• [2026/05] 🔥 Release LocateAnything — A generalist vision-language grounding model based on Eagle.
• [2025/12] 🎉 A native resolution variant of the Eagle model is adopted as the VLM backbone of GR00T-N1.6. Check out the tech blog for more details.
• [2025/10] 🔥 Release Eagle 2.5 source code.
• [2025/09] 🔥 Eagle 2.5 is accepted to NeurIPS 2025.
• [2025/09] 🎉 Eagle 2 is supported in Torch-TRT.
• [2025/07] 🎉 Release Eagle 2.5 model.
• [2025/06] 🔥 Eagle 2.5 is adopted as the VLM backbone of GR00T-N1.5. Check out the tech blog for more details.
• [2025/04] 🎉 Release Eagle 2.5 tech report.
• [2025/03] 🔥 Eagle 2 is adopted as the VLM backbone (System-2) of GR00T-N1. Check out the GTC launch and white paper for more details.
• [2025/01] 🎉 Release Eagle 2 tech report and models.
• [2025/01] 🎉 Eagle is accepted as ICLR 2025 Spotlight.
• [2024/08] 🎉 Release Eagle.

Resources

🌟 Get Started

• 📚 Getting started with VFlash
• 📚 Getting started with LocateAnything
• 📚 Getting started with Eagle 2.5
• 📚 README for Eagle

The Eagle VLM Family

Eagle is a family of frontier vision-language models (VLMs) from NVIDIA that explore data-centric strategies across general-purpose multimodal understanding, long-context reasoning, and embodied applications.

Beyond advancing multimodal foundation model research, Eagle has also served as a research and development platform that supports multiple flagship NVIDIA efforts across enterprise intelligence and Physical AI, including but not limited to Llama-Nemotron Nano VLM, Nemotron VLMs (V2 Nano VL / V3 Nano Omni), NeMo Retriever (Llama Nemoretriever Colembed), NVIDIA Isaac GR00T N1 (Tech Report) / N1.5 / N1.6, and Cosmos.

Model	Features	Summary
VFlash	Fast, accurate visual planning	Flash Primitive Decoding for trajectories, boxes, points, and dense motion
LocateAnything	Generalist grounding, detection and pointing	Fast & high quality vision-language grounding with Parallel Box Decoding
Eagle 2.5	Frontier VLM with SOTA image & video understanding	Framework & data strategy for long-context multimodal understanding
Eagle 2	Frontier VLM with SOTA image understanding	Exploring the post-training data strategies for frontier VLMs
Eagle	VLMs with mixture-of-encoders	Exploring the design space for vision-centric VLMs

VFlash

VFlash is a generalist decoding framework for fast, accurate visual planning in Physical AI. It uses the predictable structure of visual primitives to generate complete plans in parallel. A lightweight drafter learns from annotated geometry and reads the target VLM's visual and language context to predict trajectories, boxes, points, and dense 3D motion.

• ⚡ Flash Primitive Decoding — generate a complete visual primitive in one parallel drafter pass.
• 🧩 Generalist visual planning — support robot and driving trajectories, boxes, points, dense motion, and sub-task text.
• 🎯 Higher trajectory accuracy — 18–61% lower 3D robot trajectory ADE than the autoregressive target.
• 🚀 Faster robot inference — 5.1–5.2× lower per-step MolmoSpaces latency on Jetson Thor and DGX Spark.

See the VFlash README for code, training, and evaluation.

For VFlash visual planning, downstream applications, and latency results, see VFlash Detailed Results.

Features and Capabilities

Generalist Grounding & Understanding

Dense Object Detection LocateAnything performs diverse localization tasks under a unified VLM — document understanding, GUI grounding, dense object detection, and OCR.

Fast Decoding Speed Parallel Box Decoding (PBD) vs. Quantized Coordinate Decoding — PBD predicts each bounding box atomically in a single forward pass for substantially faster throughput.

Long-Context Multimodal Understanding & Reasoning

User Prompt Analyze the video and divide it into distinct sections. For each section, create a title using a few words or phrases, and a detailed caption describing the content of the section, and indicate when it begins in seconds. Multiple sections are separated by line breaks..

Applications across Virtual and Physical Worlds

Question: Based on this graph, how did the population of Puerto Rico change over the last decade? Answer: The graph illustrates the population growth of each state in the United States between 2010 and 2020. The majority of states experienced an increase, with only two experiencing a decline. Puerto Rico is among these two states. The number of people living in Puerto Rico decreased by approximately 365,000 over the last decade. This represents a decrease of about 12%.

Examples of 3D object detection in the wild by LocateAnything3D.

Model Zoo

📦 LocateAnything Models

Model Name	Date	LLM Backbone	Vision Encoder	Max Length	Download
LocateAnything-3B	2026.05.26	Qwen2.5-3B-Instruct	MoonViT-SO-400M	25K	🤗 HF

📦 Eagle 2.5 Models

Model Name	Date	LLM Backbone	Vision Encoder	Max Length	Download
Eagle2.5-8B	2025.04.16	Qwen2.5-7B-Instruct	SigLIP2	128K	🤗 HF

📦 Eagle 2 Models

Model Name	Date	LLM Backbone	Vision Encoder	Max Length	Download
Eagle2-1B	2025.01.11	Qwen2.5-0.5B-Instruct	SigLIP	16K	🤗 HF
Eagle2-2B	2025.01.11	Qwen2.5-1.5B-Instruct	SigLIP	16K	🤗 HF
Eagle2-9B	2025.01.11	Qwen2.5-7B-Instruct	SigLIP + ConvNext	16K	🤗 HF

📦 Eagle Models

Model Name	Date	LLM Backbone	Vision Encoder	Download
Eagle-X4-8B-Plus	2024.09.16	Llama-3-8B-Instruct	CLIP+ConvNeXt+EVA+Pix2Str	🤗 HF
Eagle-X4-13B-Plus	2024.09.16	vicuna-13b-v1.5	CLIP+ConvNeXt+EVA+Pix2Str	🤗 HF
Eagle-X5-34B-Plus	2024.09.16	Yi-34B	CLIP+ConvNeXt+EVA+Pix2Str+SAM	🤗 HF
Eagle-X5-7B	2024.09.16	vicuna-7b-v1.5	CLIP+ConvNeXt+EVA+Pix2Str+SAM	🤗 HF
Eagle-X5-13B	2024.09.16	vicuna-13b-v1.5	CLIP+ConvNeXt+EVA+Pix2Str+SAM	🤗 HF
Eagle-X5-13B-Chat	2024.09.16	vicuna-13b-v1.5	CLIP+ConvNeXt+EVA+Pix2Str+SAM	🤗 HF

Citation

If you find this project useful, please consider citing our works:

@article{chou2026vflash,
  title   = {{VFlash}: Fast, Accurate Visual Planning via Flash Primitive Decoding},
  author  = {Chou, Yu-Cheng and Wang, Shihao and Huang, Jin and Huang, Chi-Pin and Yang, Chiao-An and Fan, Xianzhe and Wang, Jing and Li, Zhiqi and Kuang, Yuanguo and Zheng, Ruijie and Bjorck, Johan and Liu, Sifei and Su, Hang and Badki, Abhishek and Zhang, Lei and Yuille, Alan and Fan, Linxi and Kautz, Jan and Yu, Zhiding},
  journal = {arXiv:2611.08449},
  year    = {2026}
}

@inproceedings{wang2025locateanything,
    title={{LocateAnything}: Fast and High-Quality Vision-Language Grounding with Parallel Box Decoding},
    author={Shihao Wang and Shilong Liu and Yuanguo Kuang and Xinyu Wei and Yangzhou Liu and Zhiqi Li and Yunze Man and Guo Chen and Andrew Tao and Guilin Liu and Jan Kautz and Lei Zhang and Zhiding Yu},
    booktitle={ECCV},
    year={2026}
}

@inproceedings{man2025locateanything3d,
    title   = {{LocateAnything3D}: Vision-Language 3D Detection with Chain-of-Sight},
    author  = {Yunze Man and Shihao Wang and Guowen Zhang and Johan Bjorck and Zhiqi Li and Liang-Yan Gui and Jim Fan and Jan Kautz and Yu-Xiong Wang and Zhiding Yu},
    booktitle = {CVPR},
    year    = {2026},
}

@inproceedings{chen2025eagle2.5,
    title={Eagle 2.5: Boosting Long-Context Post-Training for Frontier Vision-Language Models},
    author={Guo Chen and Zhiqi Li and Shihao Wang and Jindong Jiang and Yicheng Liu and Lidong Lu and De-An Huang and Wonmin Byeon and Matthieu Le and Max Ehrlich and Tong Lu and Limin Wang and Bryan Catanzaro and Jan Kautz and Andrew Tao and Zhiding Yu and Guilin Liu},
    booktitle={NeurIPS},
    year={2025}
}

@article{li2025eagle2,
    title={Eagle 2: Building Post-Training Data Strategies from Scratch for Frontier Vision-Language Models}, 
    author={Zhiqi Li and Guo Chen and Shilong Liu and Shihao Wang and Vibashan VS and Yishen Ji and Shiyi Lan and Hao Zhang and Yilin Zhao and Subhashree Radhakrishnan and Nadine Chang and Karan Sapra and Amala Sanjay Deshmukh and Tuomas Rintamaki and Matthieu Le and Ilia Karmanov and Lukas Voegtle and Philipp Fischer and De-An Huang and Timo Roman and Tong Lu and Jose M. Alvarez and Bryan Catanzaro and Jan Kautz and Andrew Tao and Guilin Liu and Zhiding Yu},
    journal={arXiv:2501.14818},
    year={2025}
}

@inproceedings{shi2025eagle,
    title = {Eagle: Exploring The Design Space for Multimodal LLMs with Mixture of Encoders}, 
    author={Min Shi and Fuxiao Liu and Shihao Wang and Shijia Liao and Subhashree Radhakrishnan and De-An Huang and Hongxu Yin and Karan Sapra and Yaser Yacoob and Humphrey Shi and Bryan Catanzaro and Andrew Tao and Jan Kautz and Zhiding Yu and Guilin Liu},
    booktitle={ICLR},
    year={2025}
}

License/Terms of Use

• The code is released under the Apache 2.0 license as found in the LICENSE file. Portions of the code in this repo are reused and subject to their original licenses. Some files have been modified, with appropriate attribution and additional license headers added where applicable.
• The pretrained model weights are released under either the CC BY-NC 4.0 License or the NVIDIA License. The models are research preview intended for non-commercial use only.
• Eagle models are improved using Qwen.
• For code contributions to Eagle, please refer to the Contribution Guide.
• Users are reminded to ensure that their use of the dataset and model weights is in compliance with all applicable laws and regulations.

Acknowledgement

• LLaVA, LLaVA-HR and InternVL: The Eagle codebase has integrated modified components from these repositories. Many thanks for the great open-source projects.
• LMMs-Eval and VLMEvalKit: We use derivatives of these repositories for evaluation. Many thanks for the wonderful tools.
• Thanks to Cambrian, LLaVA-One-Vision, The Cauldron and many other works for the great efforts in open-sourcing data.
• The team would like to give special thanks to the NVIDIA TSE Team, including Chen Fu, Yuchao Jin, Le An, and Josh Park, for their exceptional work on the optimized TensorRT and edge deployment of Eagle.