Local AI feels slow: what should we measure first?

Slow can mean several different things in a local AI workflow. This thread is for collecting observations before choosing a fix.

If you are asking for help, separate these parts of the experience:

• Time from sending the request to the first visible output.
• Time spent generating the rest of the answer.
• Whether the first request behaves differently from later requests.
• Whether the issue appears with a short input, a long input, or both.
• What changes when other applications or requests are running.

Include your hardware, runtime and model versions, exact model identifier, settings, and a sanitized test input. Record the command or method used for each timing.

If you reply with a proposed test, change one condition at a time and explain what the result would help rule out. Please post follow-up observations, including tests that did not improve the situation.