NVlabs Eagle: four model generations in one repository and no install script

Three directories carry the code and none carries an installer

The top level holds eight entries and nothing else: `.gitattributes`, `.gitignore`, `CONTRIBUTING.md`, `Eagle/`, `Eagle2_5/`, `Embodied/`, `LICENSE` and `README.md`. There is no dependency manifest, no install script, no test directory and no workflow directory at that level, so each generation has to be set up on its own. The setup instructions are named rather than linked from a single procedure: getting started with LocateAnything points at `Embodied/README.md`, getting started with Eagle 2.5 points at `Eagle2_5/document/0.onboarding.md`, and Eagle points at `Eagle/README.md`. Two entries are only implied: the mid-generation work sits under `Eagle2_5/`, while the grounding model released in May 2026 sits under `Embodied/`, a name that gives no hint about what it contains until you open the README inside it.

The code license and the weights license are two files

The badge row at the top links two separate license files, the project `LICENSE` and `Eagle2_5/LICENSE_MODEL`, and the metadata for the repository is Apache-2.0. That distinction is the practical one for anyone planning to ship a model rather than read code, because the terms that apply to the weights are in the file inside the Eagle 2.5 directory, not in the file at the root, and the text of neither is reproduced on this page. There are also no GitHub releases here, so the versioned artifacts are the model repositories on Hugging Face under the nvidia collection, with a specific 8B checkpoint published as Eagle2.5-8B. In other words, the repository itself is a moving set of directories while the weights you actually run are pinned elsewhere, so pinning a commit hash is a reproducible way to identify the code used.

Parallel Box Decoding puts each box in one forward pass

The grounding model takes the localization work that vision-language models usually spend tokens on. The stated comparison is between Parallel Box Decoding and quantized coordinate decoding, with the claim that Parallel Box Decoding predicts each bounding box atomically in a single forward pass, which raises throughput substantially. Underneath that sits a unified model handling document understanding, GUI grounding, dense object detection and OCR in one system, so the same weights serve four jobs that would otherwise need separate detectors. The June 2026 updates are the practical part: batch inference arrived on a pure FlashAttention runtime, described as efficient on A100, RTX 4090 and other non-Hopper and non-Blackwell GPUs, a visual prompt LoRA fine-tuning script was released under `Embodied/shell/`, and the model was accepted to ECCV 2026 in the same month. A project built by NVIDIA stating that its fast path is the one for older architectures is worth reading twice before planning hardware around it.

LocateAnything is a fourth line, not a version number

The family table gives four entries and they are not four releases of one thing. LocateAnything is the grounding model for detection and pointing, Eagle 2.5 is the image and video model built around a framework and data strategy for long-context multimodal understanding, Eagle 2 explores post-training data strategies for frontier vision-language models, and the original Eagle explores the design space for vision-centric models with mixture-of-encoders. The publication trail runs alongside: the original Eagle was accepted to ICLR 2025 as a Spotlight, Eagle 2.5 to NeurIPS 2025, and LocateAnything to ECCV 2026. The naming history is visible in the links too: the October 2025 entry for Eagle 2.5 source points at a path spelled `NVlabs/EAGLE`, while the repository itself is `NVlabs/Eagle`, so one line of the update log disagrees with every other reference to the same place.

A native resolution variant shipped downstream only

The most consequential entry in the update log is a deployment rather than a release. In December 2025 a native resolution variant of the Eagle model was adopted as the vision-language backbone of GR00T-N1.6, and the earlier entries record Eagle 2 as the System-2 backbone of GR00T-N1 in March 2025 and Eagle 2.5 as the backbone of GR00T-N1.5 in June 2025. None of the three backbones is a directory here, and the native resolution variant has no name, no path and no release of its own in this tree. The same pattern holds across the wider list of work the project says it supported: Llama-Nemotron Nano VLM, the Nemotron V2 Nano and V3 Nano Omni models, NeMo Retriever, the Isaac GR00T series and Cosmos. What runs in production downstream is mostly code that does not live in this repository.

The example prompt ends with two periods

The long-context section is written as a demonstration rather than a specification. The sample user prompt asks the model to analyze a video, divide it into distinct sections, give each one a short title and a detailed caption, and say when it begins in seconds, and it closes with a doubled period after the phrase about line breaks. The sample output answers with a first section labelled at 1.05 seconds, meaning the segmentation shown does not start at zero, followed by a paragraph of scene description for a warrior on a mountaintop. The rest of the visible page follows that shape: embedded videos for dense detection and decoding speed, report PDFs for Eagle, Eagle 2 and Eagle 2.5, a report for LocateAnything, links to a Hugging Face collection and a demo space, and a project page. Capabilities are illustrated rather than specified, and the accompanying text is short enough that the reports carry the actual detail.

Setup differs per generation and the last push predates the batch path

Because there is no release history and no shared installer, a working setup means picking a generation first. The original Eagle and Eagle 2 are described through their paper-style reports and model pages, Eagle 2.5 through a numbered onboarding document, and LocateAnything through its own README with a shell script under `Embodied/shell/` for LoRA fine-tuning on visual prompts. The last commit landed on 24 June 2026, in the same month the LocateAnything updates went in, and after that the pages of interest are the per-model documents rather than the root. Two release markers also fall outside the repository: Eagle 2 was reported as supported in Torch-TRT in September 2025, and Eagle was accepted to ICLR 2025 as a Spotlight, so a downstream integration can be newer than anything visible in the code.

Editorial conclusion

Eagle fits researchers reading model code and papers rather than teams looking for a pip install, since each generation carries its own onboarding document and the weights arrive from model hubs instead of releases. Three things are worth confirming before you build on it. The code is Apache-2.0 but the weights sit under a separate license file in the Eagle 2.5 directory, so read the one that matches what you intend to ship. There are no tagged releases, which means pinning a commit rather than a version. And the grounding line, LocateAnything, is the part with its own runtime story: its batch inference path targets A100, RTX 4090 and other non-Hopper and non-Blackwell GPUs, so hardware assumptions differ from the rest of the family.

Frequently asked questions

Official sources

1. Issues
2. License: Apache-2.0
3. NVlabs/Eagle on GitHub
4. Project website
5. README