NVlabs Eagle Trends at 1.4K Stars With +264 Today: NVIDIA's Frontier Vision-Language Model Ships 512-Frame Video Comprehension and Eagle-Video-110K Dataset

GitHub·medium signal

NVlabs/Eagle is NVIDIA's family of frontier vision-language models designed for long-context multimodal learning. Eagle 2.5 supports up to 512 video frames for long video comprehension (few minutes to 3 hours), trained jointly on image and video data. Ships with Eagle-Video-110K, a novel dataset with story-level and clip-level annotations. Eagle2-9B achieves state-of-the-art across multimodal benchmarks, matching models with up to 70B parameters at a fraction of the size.