With openAI's compute they likely can do the post-training for a decision model on a luna like model in the order of days. I woulnd't be surprised if the stuff before and after that training took more time.