STORM introduces a novel spatiotemporal token reduction approach that dramatically improves how multimodal LLMs process and understand long videos.

• Reduces computational cost by efficiently compressing video information
• Enables explicit temporal modeling between frames, capturing dynamic patterns
• Achieves state-of-the-art performance on long video understanding tasks
• Offers up to 16x better efficiency compared to traditional frame-by-frame processing

This engineering breakthrough addresses a critical limitation in current video LLMs, making long-form video analysis more practical and effective for real-world applications in content analysis, video search, and automated understanding.

Token-Efficient Long Video Understanding for Multimodal LLMs