Output tokens generated per second.
Throughput is the rate at which a model streams output tokens after the first token arrives, usually reported in tokens per second.
Low throughput makes long answers feel slow even when TTFT is fast. Compare measured throughput, not marketing labels.