“A common optimization strategy is to use smaller models like a 7-billion parameter Llama model for simple tasks such as classification, while reserving larger models for user-facing generation and critical control flow decisions.”