i've found llama.cpp (as i understand it, ollama now uses their own version of this) to work much better in practice, faster and much more flexible.