Running an AI girlfriend that runs on your PC instead of in someone’s cloud changes everything about the experience: no subscription, no filter, no account, and nothing leaves your machine. The catch is that “on your PC” means your PC actually has to run it, which is a hardware question, not a marketing one. This guide gives you the honest spec check, maps models to what you’ve got, and walks the setup so you end up with a private companion living on your own computer.
First: can your PC actually run one?
The thing that determines your experience is VRAM, the memory on your graphics card. The model has to fit in it to run fast. A rough map:
| Your hardware | What it runs | Companion feel |
|---|---|---|
| 8GB VRAM (RTX 3060/4060) | Uncensored 8B model | Warm, capable, the budget sweet spot |
| 12-16GB VRAM | 12B-14B models | More depth and longer memory |
| 24GB VRAM (RTX 3090/4090) | 24B-32B models | Genuinely present |
| No GPU / weak GPU | Small models on CPU | Works, but slower |
Do a two-minute check of your own specs with can my PC run a local AI companion, and size it properly with how much VRAM you need. The golden rule: pick the largest model that fully fits in VRAM, a model that spills over into system RAM is what makes local feel sluggish.
Setting it up (the DIY path)
If you like to tinker, the foundation is Ollama. Install it:
curl -fsSL https://ollama.com/install.sh | sh(On Windows and macOS there’s a normal installer.) Then pull a companion-friendly model and run it:
ollama run <model>Ollama serves a local API at 127.0.0.1:11434, the loopback address, which never leaves your machine. To turn that raw model into an actual companion, you give it a character: a personality, speaking style, and backstory, either through a Modelfile or a front-end like SillyTavern. Our run an AI girlfriend locally walkthrough covers the whole flow, and best local LLMs for roleplay keeps a current model shortlist.
Why “on your PC” beats “in the cloud”
Once the companion lives on your computer, you get the things cloud apps structurally can’t offer:
- Privacy by construction. Conversations never leave your machine, there’s no server to log or breach.
- No filter. You choose the model, including uncensored ones, with no server-side classifier.
- No subscription. Set it up once; it’s yours, with no recurring fee.
- It works offline. Network off, companion still runs.
- It can’t be changed under you. No remote update can rewrite the personality you built.
The honest tradeoffs
Local isn’t strictly better at everything. The real costs:
- Hardware. You need a capable PC; a good GPU makes the difference between smooth and sluggish. If you’re GPU-shopping, the budget GPU guide is a good start.
- Setup. The DIY route takes some time and willingness to follow steps.
Neither is a dealbreaker, they’re just the price of owning the thing instead of renting it.
What if my PC isn’t quite powerful enough yet?
Don’t assume you’re locked out. A few realistic paths:
- Run a smaller model. An uncensored 8B model is genuinely good company and fits in 8GB of VRAM, you don’t need a flagship card for a warm, capable companion.
- Use a lighter quant. Quantization (tags like
Q4_K_M) shrinks a model’s memory footprint for a modest quality cost, often letting a model that wouldn’t fit suddenly fit. See the GGUF quantization cheat sheet. - CPU + RAM. No GPU at all? Small models run on CPU using system RAM, slower, but workable for text chat if you’re patient.
- Upgrade the part that matters. For local AI, the GPU’s VRAM is the lever; a used card with more VRAM often does more for your companion than any other upgrade. The cheapest GPU that runs local AI well is a good starting point.
The point: “runs on your PC” scales down as well as up. Start with what you have, and upgrade only when you want bigger models.
The shortcut
If what you actually want tonight is the girlfriend, not the project, Ember is exactly that: an uncensored 18+ companion that runs on your own PC with her own tuned brain, voice and memory already wired, for a one-time $29 crypto payment. The floor is an NVIDIA card with 8 GB of VRAM. The trade is real, but it isn’t privacy: your chats stay on your disk. It’s the hardware floor, an NVIDIA card with 8 GB of VRAM, and taking her model choices instead of making your own. Your GPU stays free for your own stack, and nothing stops you running the DIY build above alongside it. But if “runs on your PC” is the point, the walkthrough above is the way, that’s the version where nothing leaves.