TinyLlama delivered the strongest responsiveness on the Pi, making it the most usable option for lightweight local inference.