Apple trained a large language model to efficiently understand long-form video

Apple researchers developed SlowFast-LLaVA-1.5, a more efficient video-understanding language model that outperforms larger competitors. The model uses a two-stream approach—analyzing fewer frames in detail while tracking movement across more frames—allowing it to process long videos without overwhelming its context window. Despite handling only 128 frames maximum, it sets new benchmarks for video analysis while maintaining strong image understanding capabilities.