my bet is their still trying to load a too large llm into a conversational pattern and all the intermediaries are getting in the way..