If you're satisfied with ollama, I don't think there's a point to go "lower" and use llama.cpp directly.

Unless you want to tinker.