VideoLoom-8B

VideoLoom-8B is a video large language model for joint spatial-temporal understanding, introduced in a January 2026 paper accepted to ICML 2026. Built on InternVL3-8B with SAM2 integration, it combines multi-frame temporal inputs with high-resolution keyframe inputs using SlowFast visual tokens, balancing temporal coverage against spatial precision.

It is trained with LoomData-8.7k, a human-centric video dataset with temporally grounded and spatially localized captions, and is evaluated on the LoomBench benchmark of temporal, spatial, and compositional video questions. Reported results include strong performance on referring video object segmentation and temporal grounding.

A 4B sibling based on InternVL2.5-4B ships alongside it for lighter deployments. The research code, checkpoints, annotation pipeline, and training code are all released, making it a solid reference for video grounding research.