Eagle 2.5
github.com/nvlabs/eagle- Category
- AI Tools
- Rank
- No. 77Tools index
- Pricing
- Open Source
- Platform
- web
- Type
- TOOL
- Builder
- nvlabs
- GitHub
- 3.4k stars
- Added
- Jul 2, 2026
About
Eagle 2.5 is a frontier vision-language model (VLM) from NVIDIA designed for long-context multimodal understanding, excelling at long video comprehension and high-resolution image analysis. It introduces novel training techniques like Automatic Degrade Sampling and Image Area Preservation, along with a new 110K video dataset with story- and clip-level annotations. Its 8B parameter model matches GPT-4o and much larger open-source models on Video-MME benchmarks.
What it does
Eagle 2.5 interprets images and video alongside text, with an emphasis on sustained reasoning across lengthy visual material. The repository presents it as a framework and data strategy for multimodal understanding, supported by downloadable models, reports, and source code.
Why it's ranked here
The case rests on practical breadth and credible research use. Eagle 2.5 targets both image and video understanding, appears in NVIDIA’s GR00T work, and belongs to a family spanning general perception and embodied applications. However, the supplied implementation details lean heavily toward research workflows rather than a finished service.
What's good
The project covers more than model weights. It supplies training machinery, configurable low-bit and LoRA options, a browser demo, a direct inference example, and an evaluation command built around LMMs-Eval. Image preprocessing handles padding and variable-resolution patch grids, which makes the implementation useful for studying real multimodal pipelines.
Tradeoffs
This is infrastructure for researchers comfortable with Python, PyTorch, model checkpoints, and GPU execution. The direct prediction example moves tensors to CUDA and uses half precision. Packaging is thin: the setup metadata declares no dependencies, despite imports from substantial external libraries. The supplied demo also describes limited safety measures and research-preview terms.
How to use it well
Use Eagle 2.5 when testing long-form video analysis, high-detail image reasoning, multimodal training recipes, or benchmark behavior within an NVIDIA-oriented research stack. Start from the published model collection, then use the repository’s inference, demo, training, and evaluation surfaces as references. It does not provide a hosted production API, deployment control plane, or complete dependency management.
Technical notes+
README.md identifies Eagle 2.5 as the family’s long-context image and video framework and links its model collection and report. Eagle/setup.py packages the eagle Python modules as version 0.0.1 but leaves install_requires empty. Eagle/train.py builds on Transformers, PyTorch, DeepSpeed-aware state gathering, LoRA selection, quantization arguments, and multimodal adapter saving. Eagle/train_mem.py requests FlashAttention 2. Eagle/gradio_demo.py exposes a configurable Gradio chat interface, while Eagle/predict_demo.py is a CUDA inference example. Eagle/evaluate_lmms_eval.py wraps LMMs-Eval through an argparse CLI. Eagle/eagle/mm_utils.py implements image-token insertion, padding, arbitrary-resolution grids, patch extraction, and stopping criteria.
Observed
- Code license
- Apache License 2.0
- Model license
- NVIDIA License
- Primary implementation language
- Python
- Packaging
- Setuptools package named eagle; dependency list is empty
- Interfaces
- Python package, command-line evaluation workflow, Gradio browser demo, and direct inference script
- Compute surface
- The supplied prediction example uses CUDA and float16 tensors
- Evaluation surface
- Repository includes an argparse-based LMMs-Eval wrapper
Read from README.md, Eagle/setup.py, Eagle/train.py, Eagle/train_mem.py, Eagle/gradio_demo.py, Eagle/predict_demo.py, Eagle/evaluate_lmms_eval.py, Embodied/locateanything_worker.py, Eagle/eagle/utils.py, Eagle/eagle/__init__.py, Eagle/eagle/mm_utils.py, Eagle/eagle/constants.py, Eagle/lmms_eval/utils.py.
What it can do
Analyze and answer questions about long videos
Long video file or stream → Natural language answers, summaries, or descriptions of video content
Analyze high-resolution images
High-resolution image file → Detailed natural language descriptions, answers, or structured analysis of image content
Generate story-level and clip-level video annotations
Video file → Structured annotations describing narrative arcs and individual clip-level events
Locate and ground objects within images
Image file with a natural language query or object description → Bounding boxes or spatial coordinates identifying the queried objects
Perform multimodal question answering across image and video inputs
Image or video paired with a natural language question → Natural language answer derived from visual and contextual reasoning
Train or fine-tune vision-language models using provided scripts
Custom image/video datasets and configuration parameters → Fine-tuned VLM model weights optimized for target tasks
Benchmark vision-language model performance on Video-MME tasks
Video-MME benchmark dataset → Performance scores and evaluation metrics comparable to GPT-4o baselines
Tags
Comments (0)
No comments yet
Indexed by a proprietary survey. Corrections welcome.