Do you mean fine tuning (say via LoRA) or working with an "open weight" model or are you talking about true "open source" models. If it is the latter, I struggle to understand how a circa GPT3 model requiring millions and millions to train would be a good use of resources over one of the former.