Ask HN: How are ML models persisted once they are trained?

Assuming that a model is trained once (or continually), and that, even though the actual service (i/e. one of Microsoft's ML APIs) exists on a distributed architecture and is accessed via API. I've seen examples, using Tensorflow, where a model is instantiated and trained programmatically. However, what happens to that model once the program ends? For example, a NodeJS process calls process.exit(). I'm assuming the ML model created in that process no longer exists. Therefore, where are those models stored? Using what technology? How are they accessed and reinstantiated? Or, is it simply that the process must run forever or risk needing to be trained from nothing, again? Obviously, I find this to be unlikely given the intensity of training a complex system. Sorry if my terminology is off, but explanations, links, etc. are very welcomed, thanks.