Open source obviously preferable. Custom training not needed, just cheaper inference on existing GPT-level 'intelligence'.