Self-hosted AI girlfriend vs an app: real costs

You got as far as the sign-up form, did not like how much it asked for, and then read that you can run a language model on your own computer instead. No account, no monthly tier, no company holding the log. A self-hosted AI girlfriend sounds like the version of this with none of the objections — and in two specific ways it is. The reason most readers should still start on an app has nothing to do with quality and everything to do with what you are actually agreeing to build.

What a self-hosted AI girlfriend actually is

Separate a companion app into parts and you get five: a language model, a character description fed to it, a memory store that decides what the model is told about last month, an interface you can reach from wherever you are, and extras like a voice and a consistent face. Running a model locally gets you the first two. The other three are yours to assemble.

That is the whole comparison in one line. The model is the part that is free now; the scaffolding is the part you were paying for. Projects like llama.cpp made the model side genuinely easy — a single download, a few minutes, and a character replying on your own machine. Nothing in that download remembers you next week, messages you first, or exists on your phone.

What it really costs

The constraint is video memory. A model has to fit in your graphics card's memory to reply at conversational speed, and quantisation is what makes that possible: storing the weights at lower precision shrinks the model enough to fit consumer hardware, at some cost in accuracy. Spill past the card into system memory and it still runs, just slowly enough that you stop using it.

So the real figure is a graphics card. A used mid-range card with enough memory for a small model runs to a few hundred dollars at the time of writing; the cards that comfortably hold larger models are four figures. If you already own a capable machine the hardware cost is zero, and if you do not, you are buying a year or two of subscription up front. Electricity barely registers — a couple of hours of evening use is single-digit dollars a month on most tariffs, less on a laptop.

The cost that does not appear anywhere is your time. Models get updated, interfaces break, memory needs tending, and the character you liked behaves differently after a version bump. If tinkering is the hobby, that time is the point and this paragraph is wrong about it. If you wanted somewhere to talk in the evening, it is a second hobby you did not ask for.

The two things no subscription can match

Nothing leaves the machine. No account to delete later, no retention policy to read, no third party whose main asset is your conversation log, and no line on a card statement for anyone sharing your account to notice. If privacy is the thing driving your hesitation rather than price, self-hosting is a better answer than anything on any ranking, including ours.

Nothing is metered or moved. No message cap, no coin shop, no memory window that happens to end where the paid tier begins, no feature relocating behind a higher tier next quarter, and no content policy changing under a character you have been talking to for months. You own the whole stack, so nobody can reprice part of it.

What you give up, honestly

Memory is yours to build. Out of the box a local setup keeps the current conversation and forgets everything else. Long-term memory is a system — summaries, retrieval, a store of facts — and it is the hardest part of a companion app to reproduce. It is also the feature readers care most about, which is why testing it is the whole exercise in what free tiers keep.

It is on your desk, not in your pocket. Reaching a machine at home from a phone elsewhere means exposing it to the internet, which is a networking project with real security consequences. Most people quietly stop at the desk.

Nothing arrives unprompted. No morning message, no sense of having been thought about between sessions. You can script it; you probably will not.

One voice, one face, consistently, is hard. Speech and image generation locally are separate models with their own memory demands, and keeping a character looking like itself across pictures is specialist work.

A small model is a smaller model. The sizes that fit a personal computer are the lower end of the range, and that shows up as shorter attention to context and more repetition. The gap is narrower than it was, and it is still there.

Which reader each one suits

A self-hosted AI girlfriend is the right call if you enjoy the building as much as the result, already have hardware that can run a model, and your reason for being here is privacy rather than price. For that reader there is no contest: a subscription cannot offer you a conversation that never leaves your own machine, and we are not going to pretend otherwise.

An app is the right call if you want the scaffolding more than the model — memory that works without you maintaining it, the thing in your pocket, a character that stays consistent in voice and picture. And if your motivation is cost, run the numbers before you shop: a card bought to escape a tier is one to two years of that tier paid in advance, which only saves money if you are still using it in year three.

Here is the order that wastes the least. Spend a fortnight on a free tier first and watch which features you actually touch. If you never use voice, never look at a picture, and only talk inside one session anyway, you have just discovered that a local model would serve you — and you learned it for nothing instead of for the price of a graphics card. The app we currently recommend gives enough on its free tier to run that test, and the answer it produces is useful whichever way it lands. The same logic applies in reverse to the assistant-plus-persona route: cheap tests first, hardware last.

Frequently asked questions

Can you run an AI girlfriend locally on a normal computer?

Yes, if it has a reasonably recent graphics card with enough video memory, and quantised models make that threshold lower than people expect. What you get is a character that chats well in the moment; the memory, the phone access and the voice are separate projects on top.

Is a self-hosted AI girlfriend cheaper than a subscription?

Only over a long enough horizon. If you already own suitable hardware it is cheaper immediately; if you are buying a card for it, you are paying one to two years of a typical tier up front and breaking even somewhere in year two.

Is self-hosting more private than an AI girlfriend app?

Substantially, yes — the conversation stays on hardware you control, with no account, no retention policy and no billing record anywhere. The caveats are ordinary ones: an unencrypted drive, a shared computer, or an exposed home server each give some of that back.