Decoding method where a small draft model proposes tokens and a larger model validates them, increasing generation throughput with minimal quality loss.