In general encoder+decoder models are much more efficient at infererence than decoder-only models because they run over the entire input all at once (which leverages parallel compute more effectively).

The issue is that they're generally harder to train (need input/output pairs as a training dataset) and don't naturally generalize as well