Ollama is not a good choice - https://sleepingrobots.com/dreams/stop-using-ollama/

As for oprncode, doesn't the system prompt eat too much of the context? Local models are really constraint in regards contex, and opencode AFAIR uses a 10k of it or some thing close.