A compression technique that reduces model size by representing weights with lower precision, making the model faster and cheaper to run with minimal quality loss.