Latency between receiving a request and generating the first output token, critical for user experience.