We're using a combination of our own models and OpenAI models. For our own models, we've found success with Flan-T5 and UL2 which we've further trained on our own data.