If you're putting AI stuff on a home server or a cheap VPS, the GPU question comes up fast. We keep a ranked list of 161 open-source AI apps you can self-host, and each entry records what the project's README says about GPUs. Here's how it shakes out.
113 run without one. 31 say a GPU is optional, 8 need one, and 9 READMEs don't say.
The GPU lives in one place: the model server. In model serving, 14 of 19 projects treat a GPU as optional (Ollama, LocalAI, llama.cpp, vLLM and friends) and 5 need one. Image and video is the other spot, with three trainers that won't run without a card.
Everything else talks to the model over an API. Chat UIs, RAG, gateways, agent platforms, memory, observability, sandboxes: almost all of them run on whatever box you've got. Voice is the split one, 6 of 11 can use a GPU if you have it.
So the easy way to plan it: pick the UI and the tools for any machine, and pick the model server for whatever card you have (or don't have).
What the list can't tell you yet is RAM. Only 11 of 161 READMEs give a minimum, so there's no RAM column. Cost estimates and tested builds aren't there either; they're on the to-do list.
The full list, with what each app needs and where it tends to break, is here if you want to keep it around: https://github.com/archestack/best-of-selfhosted-ai
Written with help from AI tools. The numbers come from the list's public data file and were checked on 2026-10-09.
Top comments (1)
how do you plan to fill the RAM column, because that one will matter more than the GPU split on a home box? a README that says "GPU optional" never tells you how much system memory the model server actually holds once it's loaded, and that is usually the thing that kills the setup on a cheap VPS. i watch it on my homelab by reading nvidia-smi per process alongside the host's free RAM, so i can tell which app is really holding the card versus which is just the model server (github.com/SikamikanikoBG/homelab-monitor). the UI-on-any-machine, model-server-on-the-card split you describe is exactly the layout that makes that per-process view useful.