SmolVLM2: New Video Language Models Released

AI Impact Summary

SmolVLM2 introduces a new family of video language models, ranging from 2.2B to 256M parameters, designed for efficient video understanding across diverse devices. The release includes pre-trained models and APIs (Python and Swift) ready for immediate use, alongside interactive demos showcasing capabilities like image and video analysis, math problem solving, and scientific question answering. This represents a significant shift towards accessible video understanding, particularly with the smaller models offering competitive performance relative to their memory footprint.