According to @ak92501 on Twitter, researchers have launched STORM, a novel text-video model that reduces video input size to just one-eighth of the usual requirement while still achieving state-of-the-art performance scores. The STORM architecture integrates mamba layers between a SigLIP vision encoder and a Qwen2-VL language model. These mamba layers effectively aggregate temporal information across video frames, allowing the model to maintain accuracy and efficiency. This development highlights significant business opportunities for companies dealing with video content processing and AI-driven video analytics, as it enables faster, more resource-efficient AI deployments without sacrificing output quality (source: @ak92501, Twitter).