I think I will finetune a 3B model in the next few days for a very specific task that currently requires a 28B model to run.