Back to Gap Map Model components / Multimodal models

Eagle

NVIDIA
restricted / Overall score: 2.7

Eagle is NVIDIA Research's family of vision-language models, built around data-centric post-training recipes. Eagle 2.5, the current release, pairs a SigLIP2 vision encoder with Qwen2.5-7B-Instruct and is trained for long-context input: up to 512 video frames, high-resolution images and multi-page documents. Eagle models also serve as the vision-language backbone of NVIDIA's GR00T robot models.

The nvidia Hub account also publishes Eagle3 checkpoints; those are speculative-decoding draft models for other LLMs, a different project, and are not part of this entry.

Openness

2 high confidence
2.0
weights
open (Eagle2.5-8B safetensors on the Hub behind an automatically approved access agreement)
data
described (open-source video and document sets plus the self-curated Eagle-Video-110K, which is not released)
code
partial (stage-2 training scripts and documentation; the data recipe ships only as a debug example)
license
NVIDIA-License (Eagle 2.5 weights; non-commercial academic or non-profit research only)
eagle-2-weights
CC-BY-NC-4.0 (Eagle2-1B, 2B and 9B, the earlier release)
repository-code
Apache-2.0 (repository code)

Eagle 2.5's weights carry an NVIDIA license that allows only non-commercial academic or non-profit research use. The training code is Apache-2.0, and the earlier Eagle 2 models are CC BY-NC 4.0, so no model in the line is cleared for commercial use.

Adoption

2 high confidence
2.0

Hugging Face downloads over the trailing 30 days, summed across Eagle2.5-8B and the three Eagle 2 sizes, most of them for Eagle2-2B and Eagle2.5-8B.

Capability

3 high confidence
3.0

Eagle 2.5 reads documents, charts and long videos, level with MiniCPM-V. It does not act in a live interface or take audio input.

Verified 2026-11-12