A slow answer has three plausible causes: loading the model, reading the input, or generating the reply. Measure each before changing models.
Run a cold and warm comparison
Use the same model and prompt twice with streaming disabled. Keep the model loaded between requests. Record load_duration, prompt_eval_duration, eval_duration and eval_count from the final responses. Compare the phases rather than treating the overall time as a single metric.
Change one variable and verify
Suppose loading accounts for most of the wait on the first request and nearly disappears on the second. Keeping the model resident may help repeated work, provided the machine has memory for it. If prompt evaluation dominates, trim irrelevant text and ask the same question again. Watch output length as well: an answer can arrive sooner simply because it contains fewer tokens. Repeat the comparison and check that the shorter or faster response still answers the question. These are illustrative tests.
For a worked example, imagine a request takes 9 seconds. The response shows 6.5 seconds of loading, 1.4 seconds of prompt evaluation and 1.0 second of generation. On the next request, loading drops to 0.1 seconds and the other timings stay close. Test a longer keep_alive next, then wait the actual gap between requests and retest. If other apps slow down while the model stays resident, factor that cost into the decision.
If both requests show little loading but prompt time is high, compare the number of prompt tokens as well as the elapsed time. Repeating an identical prompt may reuse cached tokens, so record prompt_eval_cached_count when present. To test prompt length fairly, use fresh, comparable inputs that contain the same facts. If generation dominates, record tokens per second as eval_count divided by eval_duration in seconds, then compare equally useful answers. A single quick response is a weak basis for a change. Use several runs and keep timing and correctness side by side.