with some additional coding on top, smaller models like GPT2 and BERT can produce decent summaries, a few summaries per minute depending on your target quality.