Personal.AI.Computer.mov
Open-source guides to build a machine that runs open models on hardware no one can switch off — every part, every bracket, every BIOS setting, every assembly photo. Pick the size that fits your budget and your work.
Why this exists. The best models live in someone else's cloud, behind someone else's terms and someone else's government. A model you rent can be cut off overnight. A model running in your own house can't. Build it once; own it for good.
One Personal AI Computer, three configurations — from a home rig to an on-prem business cluster. Each is a complete, self-contained guide: bill of materials, 3D-printable/CNC housing files, wiring, BIOS, and assembly photos.
| Config | GPUs | VRAM | Best for | Platform | Guide |
|---|---|---|---|---|---|
| 2× | 2× RTX 5090 | 64 GB | Home — your first local rig, runs a real model tonight | Intel Xeon W5 · ASUS W790 | → 2× config |
| 4× | 4× RTX PRO 6000 Blackwell | 384 GB | Team — big models + agents running all day | AMD EPYC 9124 · ASRock Rack | → 4× config |
| 8× | 8× RTX 4090 / 5090 | 192–256 GB | On-prem business — a company's AI on its own floor; IP and data never leave | Dual AMD EPYC 9004 (Genoa) | → 8× config |
VRAM is the constraint that decides which open models fit. Rough guide (exact fit depends on quantization — model specifics live in /software):
| Build | VRAM | Open models it can serve (e.g.) |
|---|---|---|
| 2× RTX 5090 | 64 GB | Llama 70B / Qwen 72B (quantized), 30B-class at full precision, coding + agent models |
| 4× RTX PRO 6000 | 384 GB | DeepSeek-V3 / R1, Qwen 235B, Llama 405B (quantized) — frontier-class open weights |
| 8× RTX 4090/5090 | 192–256 GB | Multiple large models served at once, plus headroom for fine-tuning |
- Pick a configuration above by budget and the models you want to run.
- Source the parts from that config's Bill of Materials.
- Make the housing — print the STL files or CNC the STEP files in that config's folder.
- Assemble — follow the config's photo-by-photo assembly guide.
- Set up the software — drivers, BIOS, and serving open models locally:
/software. - Run your intelligence — point your agent at
localhostand never get cut off again.
- 2× — Home · 2× RTX 5090, 64 GB
- 4× — Team · 4× RTX PRO 6000 Blackwell, 384 GB
- 8× — On-prem business · 8× RTX 4090/5090, 192–256 GB
Bringing the box to life — OS, NVIDIA drivers, BIOS tuning, serving open models (Ollama / vLLM / llama.cpp), and connecting an agent: /software. (Setup and testing are documented today; local model-serving guides are in progress.)
Built one? Improved a part? Found a better component? See CONTRIBUTING.md — and share your build. The best community builds get featured.
Open source under the MIT License. Fork it, change it, build your own and sell it — we just want it built.
Questions? Open an issue.