Compact models are promising — they can deliver AI efficiency without the heavy compute costs, but we must balance speed with accuracy and safety