Curtis Stone, Freedom Farmers
My exact local AI build.
Everything I run, with links. It’s three layers: the machine, the software, and the knowledge you give it. Then the power that keeps it all running. If $9,000 is more than you want to spend, there’s a cheaper version.
Layer 1
The machine
The graphics card’s memory decides everything. It sets how big a model you can run and how fast it answers. Start there and build around it.
- Graphics card
- MSI GeForce RTX 5090 32G Ventus 3X OCThe part that does the thinking. Every RTX 5090 has 32 GB of memory, so any brand works.See RTX 5090 cards on Amazon
- Memory
- XPG Lancer Blade RGB, 32 GB (2 × 16 GB) DDR5-6000Or any 2 × 16 GB DDR5-6000 kit.Buy on Amazon
- Storage
- Kingston NV3 2 TB NVMeHolds the models, the offline library and your own files.Buy on Amazon
- The rest
- Power supply, case and CPU coolerYour choice. Get a 1000 to 1200 W ATX 3.1 power supply: the RTX 5090 draws up to 575 W on its own.
Expect $9,000 to $10,000 at today’s prices, most of it the graphics card. After that it costs virtually nothing ongoing.
Want to spend less? Versions from about $2,000
I haven’t built this one. It’s from a price comparison I had done, for running the same software at a fraction of the price. The trade-off: the graphics card has 16 GB of memory instead of 32, so it runs smaller models and answers more slowly.
Whichever way you go, check that the card says RTX 5060 Ti 16 GB. There’s an 8 GB card with the same name, and it’s too small for this.
Buy it built: about $2,600
- Computer
- ASUS ROG GM700Ryzen 7 8700F, 32 GB memory, RTX 5060 Ti 16 GB, 1 TB drive. About $2,300 new at Best Buy. Open-box units are often cheaper.See it at Best Buy
- Extra storage
- Kingston NV3 2 TB NVMeThe same drive as mine, added as a second drive for models and files. The 1 TB inside fills up fast.Buy on Amazon
Build it yourself: about $2,050
- Graphics card
- PNY GeForce RTX 5060 Ti 16 GBAny RTX 5060 Ti 16 GB works.Buy on Amazon
- Processor
- AMD Ryzen 5 9600XIt doesn’t come with a cooler. Get the one below.Buy on Amazon
- Memory
- Patriot Viper Venom 32 GB (2 × 16 GB) DDR5-6000Or any 2 × 16 GB DDR5-6000 kit.Buy on Amazon
- Storage
- Kingston NV3 2 TB NVMeThe same drive as mine.Buy on Amazon
- Power supply
- MSI MAG A750GL, 750 W GoldPlenty for this card, with room to spare.Buy on Amazon
Building it yourself costs a bit less than the prebuilt, and every part is standard, so you can swap any of them later.
Two other routes
- More graphics memory
- A used RTX 3090, about $2,900 all inThe build-it-yourself list with a used RTX 3090 in place of the 5060 Ti. It has 24 GB, enough to run the 27B models below fully on the card. Used ones sell for around $1,600 on eBay, with no warranty, and the card draws about twice the power.
- A Mac
- Mac mini, M5 Pro, 64 GB, 512 GB storage: $2,699Its 64 GB of shared memory holds bigger models than the cards on this page, and it’s small, quiet and low power. But a Mac doesn’t run NVIDIA’s CUDA, so image generation is much slower and some tools won’t run. The memory can’t be added later, so buy 64 GB up front.See it at Apple
Prices as of November 5, 2026. Memory and graphics cards are expensive right now and change week to week.
Layer 2
The software
All free. The model is the brain. The agent framework is what makes it work for you instead of just chatting.
- Operating system
- Linux MintFree, and you can see everything it does.linuxmint.com
- Model runner
- OllamaDownloads and runs AI models on your own machine.ollama.com
- Models I run
-
General: qwen3.6:27b, run at temperature 0.3 with a 64K context
Vision: qwen3.8:27b, to describe images
Search: nomic-embed-text, so it can search your files
Uncensored: Qwen3.8-27B heretic-ara, Qwen3.8 with its refusals removed. On Hugging Face
The first three are free in Ollama’s library. The uncensored one is free on Hugging Face. Each runs on the RTX 5090 above. A 27B model is about 17 GB, and about 20 GB in use with a 64K context.On the cheaper version’s 16 GB card, the 27B models still run, but much more slowly, because part of the model spills into system memory. Smaller models, around 14B, run well on it.
- Agents
- Hermes AgentGives the AI memory, saved skills, tools and scheduling, and lets you run separate agents with separate boundaries.Hermes Agent on GitHub
Layer 3
The knowledge
The AI is only as good as what you feed it. Two kinds of knowledge go in.
- The library
- Project NOMADA free offline knowledge server: all of Wikipedia, medicine, survival guides, Khan Academy, farming, repair, computing and maps, in one place, with no internet.projectnomad.us
- Your own archive
- What only you haveYour emails, your notes, your documents, the manuals for your equipment. Public knowledge is everywhere. Your own history is in one place.How: download your email over IMAP and save each message as a plain text file. For videos, pull the captions with yt-dlp, one text file per video.
Power
Keeping it running
- Main power
- Solar and batteryNo utility required.
- Backup
- EcoFlow Delta 3 Max Plus2,048 Wh. The AI machine runs through it, with surge protection: about 3 hours at full load, and an estimated 11 to 13 hours in outage mode.Buy on Amazon
- Outage mode
- Automatic low powerA small script watches the battery. When the power goes out, it unloads the models and turns down the graphics card, so the machine draws about a quarter of the power (about 150 W instead of 600 W). When power comes back, it switches back.
The cheaper version draws about 250 to 300 W under load, so the same battery runs it much longer.
Disclaimer
This isn’t one click. You need a capable machine, a willingness to learn, and patience for the setup.
You don’t need to be a programmer, pay for subscriptions, or ask anyone’s permission. The learning curve is a one-time cost, and what you get is yours.
Questions, requests or ideas? Email [email protected]
As an Amazon Associate I earn from qualifying purchases. It doesn’t change my recommendations or the price you pay.