Is it using the correct prompt format for the different models? You should show exactly the string that was sent to the LLM.