llama.cpp is not the same thing as ollama. It does a similar thing, but better imo. can multiplex/multithread sessions/conversations with llms (where ollama and lmstudio etc. have to queue them up), stuff like that.

I checked, ollama does not use npu

It's not a specific model, its a harness for running models. You can find ethically trained models to run on it though iirc.