Requirements
wanna.sh runs on most modern computers, but AI models are demanding. The more graphics memory you have, the bigger and faster the models you can use. The app rates every model for your computer before you download it, so you never have to guess.
Operating system
Section titled “Operating system”| System | Status |
|---|---|
| Linux (x86-64) | Supported. |
| Windows 10 or 11 (64-bit) | Supported, with the Windows app. Also runs inside WSL2 for people who prefer the command line. |
| macOS | The installer and app run. Chat models use Apple’s Metal graphics through llama.cpp. Picture, video, voice, transcription and music models run on the processor, which is much slower. |
Graphics card
Section titled “Graphics card”An NVIDIA graphics card is strongly recommended. Without one, everything still works but runs on the processor, which can be 10 to 50 times slower.
Graphics memory (VRAM) decides which models fit. For chat models, a rule of thumb: a model compressed to 4 bits needs a little over half a gigabyte per billion parameters, plus some memory for the conversation. So an 8B model fits on an 8 GB card, and a 27B model needs roughly 16 to 20 GB.
wanna.sh can share one model across several graphics cards, so their memory adds up. For example, on a computer with an RTX 3070 (8 GB) and an RTX 3060 (12 GB), a 27B chat model runs across both cards at about 20 tokens (roughly 15 words) a second, and a 3B model at about 60.
You don’t have to work any of this out yourself: Get models rates each model for your computer.
Keep your NVIDIA driver up to date. After installing, the system check in Settings tells you whether the AI libraries can use your graphics card.
Memory and disk
Section titled “Memory and disk”- System memory (RAM): 16 GB or more is recommended. Models that don’t fully fit on the graphics card spill into system memory.
- Disk space: about 10 GB for the app and its AI libraries, plus space for models. Document search needs a small search model (about 640 MB) and room for your documents. Small models are 1 to 5 GB; big ones are 15 to 30 GB or more each.
Software
Section titled “Software”The Windows app needs nothing else. It sets up everything it uses by itself.
On Linux and macOS, the install command checks for these and tells you what’s missing:
- git
- Node.js 20.9 or newer, to build the app’s interface. Get it from nodejs.org or with nvm.
- curl
The installer sets up Python itself, using uv.
Optional: Docker is needed only for Code, which runs the agent’s commands inside a Docker container. On Windows, use Docker Desktop (with WSL integration turned on if you run wanna.sh in WSL2).
Internet
Section titled “Internet”You need a connection to install wanna.sh and download models. After that, models work offline.