For your use case, I recommend the GPT-2 model. Fast, cheap, and open weights!