In reply to Christine Lin:
One of the significant benefits of using UVM is that you can run all of your tests with different random seeds without re-compiling. When you run all of the tests, the DUT should be the same so that the coverage metrics can be merged and managed correctly.
I would recommend using separate compilations to different libraries, one for each DUT model. You can then run regressions against each compilation. I would not recommend changing DUT models in the UVM environment since this will cause merging issues when determining coverage.