Models aren't just big bags of floats you imagine them to be. Those bags are there, but there's a whole layer of runtimes, caches, timers, load balancers, classifiers/sanitizers, etc. around them, all of which have tunable parameters that affect the user-perceptible output.