How would you optimize LLM performance, manage context windows, and improve fine-tuning cost efficiency for Mistral AI?