would probably need multiple models running in distinct containers, with another process coordinating them