Generation is a sequence of choices made from a learned sense of what is plausible. Because more than one option can be plausible, and because the choice involves randomness, the output is not fixed in advance.
Why two runs differ
Ask a text tool for a name for a coffee shop. The first run returns "The Daily Grind." The second run, same prompt, returns "Morning Ember." Both fit the request. Neither was stored anywhere. The model sampled a different plausible option the second time.
If the model were retrieving stored answers, repeating a prompt would return the identical result every time. It often does not, which is one everyday sign that sampling is happening.