Merging model outputs also means you have to run both models. I think the best option is to merge the weights and recover performance using datasets from both domains and distillation from the respective expert model.