This paper advocates for a shift from model scaling to data-centric efficiency, demonstrating that for edge-based video AI, a 3B parameter model is often superior to a 7B model. The researchers discovered an "edge-scaling paradox" where the smaller 3B model matches the 7B in accuracy while delivering significantly higher throughput and lower memory strain. By replacing traditional "black-bar" padding with Naive Warping (stretching video to fit the model), they eliminated "dead" tokens and halved the memory footprint without any loss in reasoning performance. Ultimately, the study proves that when hardware is limited, a highly compressed, smaller LLM—focused on dense, high-utility visual data—is the most stable and efficient choice for real-time deployment.