A post this morning asked the question plainly: what should I download and back up right now, given that NVIDIA is buying Hugging Face and the future feels uncertain.

We have some responsibility for that mood, because a week ago we wrote about the acquisition and said that if your work depends on a specific model you should keep your own copy of it. Then four days later we wrote about our own disk filling up and taking our publishing offline, next to somebody who had just deleted 3.2 TB of models in one go.

Both of those pieces are right, and together they are useless advice. So here is the reconciliation.

Nothing has actually happened yet

Worth saying first, because it is the part that should lower everyone’s heart rate. The deal has not closed and is not expected to until well into next year. No models have been removed. Both companies have said the platform stays open and that NVIDIA hardware will not be required to use it.

We are watching specific things rather than vibes: whether hosting stays free and generous for large files, whether non NVIDIA hardware keeps working as well as it does today, and whether the number of models sitting behind an access request climbs. None of those have moved. If they move we will write about it.

Hoarding three terabytes today against something that has not happened, and might not, is how you end up with our problem instead of theirs.

The four things that genuinely deserve a copy

Anything you cannot reproduce. Your own trained weights, your own datasets, the LoRA you spent a weekend on. These are irreplaceable in the literal sense and they are usually the smallest files you own. We checked our own while writing this and found 50 GB of trained weights sitting on exactly one drive, which is the drive that filled up on us last week. So take this one as advice we are about to follow rather than advice we had already taken.

The exact version you depend on. Not the model, the version. If a workflow you rely on works against one specific revision, that revision is your dependency, and “it is on Hugging Face” is not the same as “that exact file is on Hugging Face next year”. Pin the revision rather than the repository name.

Anything already living on borrowed time. Models with licences that could be withdrawn, models from small teams with no institutional backing, anything that has been reuploaded by a third party because the original vanished once already. The pattern to watch is not corporate acquisition, it is abandonment.

Things that are already hard to find. If it took you an afternoon to locate, it will take you another afternoon, and the thing you cannot get back is the afternoon.

What does not deserve it

Almost everything else, honestly.

The big, popular, actively maintained checkpoints are the least likely things on the internet to disappear. They are mirrored, torrented, cached and reuploaded within hours of any hiccup. They are also the files most likely to be superseded by something better before you next open them, which is the part the hoarding instinct ignores. Most of a large model collection is not insurance, it is a museum of things you were briefly excited about.

Our own ComfyUI models folder is 719 GB. If we are honest, the share of that we would actually miss is small, and the share we could redownload in an evening is most of it.

One more thing that is easy to miss: a copy on the same machine is not a backup. It is a second copy of a file that will die with the same drive. Our disk filled up, and everything on it was equally unavailable regardless of how carefully we had collected it.

The unglamorous version

Work out which two or three models your actual work depends on. Pin their exact revisions. Put those, plus anything you trained yourself, somewhere that is not the drive you render on. Then stop.

That takes an hour and it protects you against every scenario that is realistically coming, including ones nobody is worried about today. The alternative, downloading everything you might one day want, protects you against very little and guarantees the failure mode we hit last week.

There was also a nice counterpoint on the radar this morning. Somebody trained a 210M parameter image model from scratch on a single GPU in three and a half days. Not a fine tune, a model. It is small and it is not competing with anything you use, but it is a reminder that the ability to make these things is spreading rather than concentrating, which is a better hedge than a full hard drive.

A daily note from our news radar. The week gets the full treatment in the DIY AI Brief every Monday.