Technically? Not a problem. The reason most of them run in data centers is the massive amount of computational and therefore also electrical power you need to run a somewhat useful model.
Even worse when you need to initially train them. That'll really hit the wallet.
There are (by now) vast selection of models you can easily run at home without any outside connection, as long as you have reasonable hardware to run them.
whereas the other options basically force you to forever use their database-based system and files are terribly organized, so you are forced to use their interface.
Immich has a "storage templates" section which allows you to choose a folder structure that it will use to store the files in.
Or go the other way and include your folders as external libraries.
Yeah, that's not going to work. You'll have to at least rank your requirements. Is size more important, or is it the number of RAM slots? Also, what is "enough CPU/Ram"?
Also also, "unlimited" will only get you unrealistic things like the 10k+ PCI-E SSDs.
It grew from a nice Owncloud fork into a do-it-all groupware solution by adding on more and more things without really improving the basis. Each version the performance gets a little worse, syncing gets stuck more often, etc.
Opencloud looks or at least looked good as it started out as an Owncloud Infinite Scale fork, but of course they're adding on more and more groupware stuff without improving the core first. Maybe we're doomed to witness the same cycle with each solution, who knows.
You didn't put a country, but wherever you are: check the laws regarding underage relationships, if any. He might actually get into trouble depending on what they're up to.
They all will do that, some more, some less. It's to stop them from sticking in position.