README says that the git backend is the recommended backend, as the "native" one has no additional features, so I imagine: it handles them the same as git (ie. they are just objects in the .git repo data, and each time you change them you add a new one, and they are poorly compressible and optimizable) -- which is, I imagine, the problem you're referring to.

It won't be completely the same as git, because the client fetches stuff more lazily (according to the linked presentation). So the backend will still fill up with stuff and need lots of storage, but the clients then won't necessarily slow down in the same way.

That said, any git users will still get slowed down - only jj users would see the benefit. Git does have better features these days for shallow clones though and even git is on the way to killing off the need for LFS.

Might want to look at purpose built tools for that such as lakeFS (https://github.com/treeverse/lakeFS/)

* Disclaimer: I'm one of the creators/maintainers of the project.

Now I'm curious: how would you want a DVCS to handle large binaries?

Partial checkout/Shallow first checkout.

JJ doesn't seem to do that though, right?

Probably something block based, such as borg