The ability to automatically adjust how many of a model's parameters are actively used based on available computational resources, allowing the same model to run efficiently on different hardware.