What your computer needs
The interface is light; the model is what needs power. SillyTavern "will run on anything that can run NodeJS 20 or higher," but for local models its README recommends "a 3000-series NVIDIA graphics card with at least 6GB of VRAM" (SillyTavern README).
- Memory: KoboldCpp's wiki says a 7B model needs at least 8 GB of RAM and a 13B model at least 16 GB, with a Q4_0 quantized model and a 2,048-token context, and that offloading layers to the graphics card reduces that (KoboldCpp wiki).
- Disk: Ollama's Llama 3.1 8B download is 4.9 GB (Ollama) and Mistral Nemo 12B is 7.1 GB (Ollama); SillyTavern's docs say local models "can be 5-50GB each" (SillyTavern docs).
- Graphics: Ollama supports NVIDIA cards with compute capability 5.0 or newer and driver 550 or later (Ollama docs).
- Mac: LM Studio needs Apple Silicon and recommends 16 GB of RAM or more; Intel Macs are not supported (LM Studio).
In practice, an 8B to 12B model is where roleplay starts to feel coherent, and that means a recent gaming PC or an Apple Silicon Mac with 16 GB or more.
A basic local setup in five steps
- Install a backend. KoboldCpp is a single download; Ollama installs like a normal app and pulls models by name.
- Download a model. Start with an 8B to 12B model in a 4-bit quantized version that fits your graphics memory. KoboldCpp's README even names a roleplay model to start with.
- Install SillyTavern. It needs Node.js 20 or newer and runs in your browser at a local address.
- Connect the two. In SillyTavern's API connections, choose your backend and point it at the local address the backend shows.
- Create her. Write a character card with a description, personality and first message, or import one. Add lorebook entries for facts she should remember.
Expect an evening of setup and some trial and error with models and settings. Our AI girlfriend prompts guide has character descriptions you can adapt for a card.
3D and voice companions on GitHub
Searches for "3D AI girlfriend GitHub" and "AI waifu GitHub" lead to projects that add a body and a voice to the chat. Open-LLM-VTuber calls itself a "voice-interactive AI companion" with a Live2D avatar and says "all functionalities can run completely offline on your computer" (Open-LLM-VTuber). AIRI describes itself as a self-hosted companion and runs on the web, macOS and Windows (AIRI). Amica talks through 3D VRM characters with speech recognition and connects to Ollama or llama.cpp (Amica).
These add work on top of the basic setup: speech-to-text and text-to-speech models need their own memory, and each project has its own install guide. They are the most private way to get a talking companion, and the most demanding.
Local vs hosted: what you gain and what you give up
Running your own AI girlfriend vs using a hosted app
| What matters |
Local open-source setup |
Hosted AI girlfriend app |
| Privacy |
Chats stay on your computer if the model runs locally |
Chats are stored on the company's servers |
| Cost |
Free software, but you supply the hardware |
Subscription; the median cheapest plan we track is $7.75 a month |
| Setup |
Hours, plus ongoing tweaking |
Minutes |
| Pictures and video |
Needs separate image models and more graphics memory |
Built in on most apps |
| Voice calls |
Possible with extra projects |
Built in on 15 apps we track |
| Memory |
Lorebooks and summaries you manage |
Automatic, editable on a few apps |
| On your phone |
Hard; the model usually runs on a computer |
Works in any phone browser |
One catch on privacy: if you connect SillyTavern to a cloud API instead of a local model, your messages go to that provider, and the privacy advantage disappears. For a comparison of how hosted apps remember you, see our AI girlfriend with memory ranking.
Licenses, bots and responsibility
Several of these projects use the AGPL-3.0 license, which requires you to share your changes if you modify the code and let other people use it over a network. Model weights come with their own licenses, separate from the software, so check both before building anything public.
Running a companion for other people, such as a Telegram or Discord bot built from a GitHub project, makes you the operator: you hold their messages and are responsible for them. Local software also does not change the law on what content is allowed. For more on the hosted side, see our guides on AI girlfriend app safety and creating your own AI girlfriend in an app.
Frequently asked questions
Is there an open-source AI girlfriend?
Yes, in the sense that you can build one from open-source parts. SillyTavern (a chat interface) with KoboldCpp or Ollama (to run a model) is the most common setup, and projects such as Open-LLM-VTuber and Amica add voice and an animated avatar.
How do I make a local AI girlfriend?
Install a backend such as KoboldCpp or Ollama, download an 8B to 12B model that fits your graphics memory, install SillyTavern, connect it to the backend and write a character card. Expect a few hours of setup.
What GPU do I need to run an AI girlfriend locally?
SillyTavern's README recommends an NVIDIA 3000-series card with at least 6 GB of video memory for local models. KoboldCpp's wiki says a 7B model needs at least 8 GB of RAM and a 13B model 16 GB, less if you offload to the graphics card.
Is SillyTavern free?
Yes. SillyTavern is free, open-source software under the AGPL-3.0 license. It is only an interface, so you also need a backend: a free local model runner, or a paid cloud API that charges for what you use.
Is a local AI girlfriend more private?
Yes, if both the interface and the model run on your computer, because nothing leaves it. If you connect the interface to a cloud AI service instead, your messages go to that service.
Can I run a local AI girlfriend on my phone?
Not well. SillyTavern's README links an Android install guide, but that runs the interface; the model usually still runs on a computer or through a cloud API. Hosted apps work in any phone browser.
