A smaller, faster version of a larger model created by training it to mimic the larger model's behavior, reducing computational requirements while maintaining reasonable performance.