The best self-hosted AI for a 4 GB NAS

Most "best self-hosted AI" lists quietly assume a server with 32 GB of RAM and a discrete GPU. This one assumes the machine sitting in your cupboard: two cores from around 2012, 4 GB of RAM, a pair of spinning disks and a gigabit port.

At that size the question is not "which model is smartest". It is "which process is allowed to hold memory while the others run". Our test box is an 8 GB, 2-core laptop board repurposed as a NAS — we cut the budget to 4 GB for this guide, because that is what most people have.

What actually fits

The four picks below are ordered by how much value you get per gigabyte. None of them needs a GPU and none of them calls home.

The rule of thumb: one memory-hungry service at a time. Immich's thumbnail queue, a document OCR batch and a resident 7B model cannot all run in 4 GB. They can all be installed; schedule them so they do not overlap.

What does not fit, and why

Anything that runs a large language model above about 3B parameters, and any image generator. A 7B model at 4-bit quantisation needs roughly 5.5 GB once you add context to the weights — that is the entire machine, twice the budget, before the operating system has an opinion. Image generation needs a GPU to be usable at all; on CPU you are measuring minutes per picture, not seconds.

One more thing worth more than any tool

Put a watchdog on the machine and back up the database, not just the files. Every tool here stores its state in Postgres or SQLite, and a photo library restored without its database is a folder of numbered files. That is the difference between a two-hour recovery and a two-week one.