Average small language model response