> it is well optimized for fast inference

do you have any insight into the actual technical details that make this sort of things possible? I want to learn more about model architectures. Does it have to do with attention mechanisms or sparsity or something?