Choose hardware for local AI at home
This page may contain Amazon affiliate links. A purchase after a click can earn the site a commission, at no extra cost to you. The advice stays independent. How affiliation works.
This guide is for households that want to run Ollama, Whisper, or ComfyUI at home, without a data-center server. A local LLM (a large language model) mostly uses RAM, VRAM (memory on the graphics card), and an SSD. The goal is to line up what you actually do, your budget, and your limits (noise, electricity, space in the living room), then open the more specific guides. The tags GGUF, Q4, and Q5 are explained in the glossary on this page.
On this page
- Step 1 — list what you will actually do
- Step 2 — a quick decision table
- Interactive hardware selector
- Checklist before you order
- Hardware FAQ
Rest of this guide: buying profiles · software after the hardware · cost and trust · glossary · a careful five-week path
Who this is for, and how long
| Profile | What you need first | Rough time | Path on this page |
|---|---|---|---|
| Beginner (first hardware purchase) | A first local AI in 15 minutes on the household PC | ~20 min | Uses, then the selector, then the checklist |
| Intermediate (a RAM or GPU upgrade) | You have already tried Ollama or LM Studio, and you know RAM or VRAM is the bottleneck | ~15 min | The decision table, then VRAM or 32 vs 64 GB of RAM |
| Images and chat | You are considering ComfyUI or FLUX | ~25 min | The image-generation row of the table, then GPU and eGPU before the selector |
| Careful (no impulse order) | You already know the budget and the kilowatt-hour price on the bill | ~5 weeks | The weekly path below |
Step 1: list what you will actually do
| Use | The resource that matters | A useful guide |
|---|---|---|
| Family chat, documents | 32 GB of RAM, a 7B–8B model | A private family assistant |
| Home automation | A mini PC or a NAS left on | Home Assistant is an open-source home hub. Home Assistant and Ollama |
| Email and PDF automations | RAM, plus Docker | n8n connects workflows between apps. n8n and a local model |
| Voice transcription | Mostly the CPU | Whisper, offline |
| AI images | 12 GB or more of VRAM | Images without a cloud |
| Hobby code | 16–32 GB | Continue, a local coding assistant |
| A first proof, with nothing new to buy | The PC you have, 16 GB or more | A first local AI in 15 minutes |
If several rows fit, start with family chat. That is the most common step before home automation or n8n. Also note who will use the machine: one adult in the evening, or several Open WebUI tabs on the weekend. That affects RAM and noise more than the first GGUF model you pick.
Step 2: a quick decision table
| Household profile | RAM | GPU / VRAM | Form | LocAIhome guide |
|---|---|---|---|---|
| Try it on the PC you have | 16–32 GB | Optional | Existing PC | 15 minutes |
| A quiet living-room server | 32 GB | iGPU, or a mid-range RTX | Mini PC | 2026 mini PC comparison |
| Apple is already in the house | 24–32 GB unified | Metal | Mac mini | Mac mini M4 for local AI |
| Storage plus services | Varies | Often the NAS CPU | NAS | Synology NAS and Ollama |
| Images and chat | 32 GB or more | RTX, 12 GB or more | Tower | GPU, iGPU, and OCuLink eGPU |
VRAM in detail: VRAM and which model fits your GPU. RAM in detail: 32 vs 64 GB for an LLM. Disk: an NVMe SSD for Ollama and GGUF models. Noise and a 24-hour bill: mini PC power and noise.
Interactive hardware selector
After step 1, the LocAIhome hardware selector lines your answers up with the table above and estimates total cost over 12 and 36 months. That is the same idea as the 3-year cost and the server kilowatt-hour budget. Have three facts ready before you open the tool:
- The uses you want (chat, home automation, transcription, images, and so on)
- The GGUF models you are aiming at (for example a 7B Q4_K_M versus a 13B) and a noise limit, in dB
- A rough hardware budget, and the price per kilowatt-hour on your bill
The tool offers three profiles: Budget, Balanced, and Comfort. Compare the result with the related guides and the checklist before you order anything. It filters. It does not replace reading the RAM and VRAM requirements.
Buying profiles and example store searches
Hardware-only ranges for October 2026. Check the current price at the store when you order. The labels in the table open an example search on amazon.co.uk through an amazon.co.uk partner link (how affiliation works). LocAIhome may earn a commission, at no extra cost to you. Always cross-check RAM and VRAM compatibility, the current price, and the noise figure on the spec sheet, at Best Buy, Micro Center, or Apple.com, depending on the product.
| Profile | Rough range | Where to look |
|---|---|---|
| Budget | €450–800 | 32 GB mini PC |
| Balanced | €800–1,300 | Ryzen, 32 GB |
| Comfort | €1,300–2,500+ | RTX or Mac |
Always compare with Best Buy, Micro Center, or Apple.com, depending on the product. The 3-year cost comparison helps you place the purchase next to a cloud subscription.
Software after the hardware
- Install Ollama on Windows, Linux, or Mac. If you are starting out, use Ollama on Windows, step by step.
- Choose a first model. The instruct and Q4_K_M tags are explained there.
- Open WebUI for beginners, or the Docker stack when several people will use it.
- Optional: LM Studio, to tune GPU layers before you settle on Ollama.
Cost and trust
- Cost over 3 years versus ChatGPT Plus
- Electricity budget for a server left on all day
- Privacy when AI stays home
Checklist before you order
- The main use is identified (chat, home automation, images, and so on).
- You know the noise limit, in dB, and the space (living room versus basement). See noise and power.
- Is electricity 24 hours a day acceptable? Estimate kilowatt-hours.
- The budget covers 36 months of total cost, set next to ChatGPT.
- You have been through the selector. Use the table above if you have not.
- The SSD is sized for the GGUF library. Storage guide.
Short glossary
| Term | Meaning |
|---|---|
| GGUF | GPT-Generated Unified Format, a quantized model file for Ollama and LM Studio. Q4 and Q5 are size-to-quality tradeoffs. |
| VRAM | GPU memory. See the VRAM guide. |
| Mini PC | A small case, often quiet. Comparison. |
| NAS | Network storage. Inference is limited. Ollama on a NAS. |
| Metal | Apple’s graphics acceleration on a Mac mini. M4 guide. |
Guides to open next
After this overview, read what you need: VRAM, RAM, SSD, GPU and eGPU, mini PC comparison, NAS, Mac mini, noise and power, Whisper, ComfyUI and FLUX.
Safety at home
Guest network, the Ollama port, and model backups are covered together in privacy when AI stays home. Family accounts are in the private family assistant.
Hardware FAQ
Can I reuse an old PC?
Yes, for a 3B–7B GGUF, if you have 16–32 GB of RAM and an NVMe SSD. Avoid a spinning hard drive when you load heavy models. Try 15 minutes first.
Do I have to pay for software subscriptions?
Not for Ollama, self-hosted Open WebUI, or open-source Whisper. You mostly pay for hardware, plus electricity if the machine stays on (kilowatt-hours times the rate on your bill).
How do I test without buying anything?
Run a first local AI in 15 minutes on the PC you have, then choose a first model before you open the selector.
An external GPU upgrade?
See GPU, iGPU, and an OCuLink eGPU if you want more demanding models, or images without a cloud, without replacing the whole case.
A mini PC left on, or a Mac mini?
For quiet and power draw, a Mac mini M4 or a fanless mini PC is the usual answer (comparison). For NVIDIA flexibility, use a mini PC or a tower (GPU).
Quantization and Ollama tags, so you do not buy twice
Before you order 64 GB “just in case,” test on the PC you have with choose a first model. A 7B instruct Q4_K_M is still the household standard in October 2026. Instruct tags are models trained to follow instructions. They are different from raw base models. Ollama exposes them in ollama pull. If you step up to a 13B, check 32 vs 64 GB of RAM and VRAM. That is often a better use of money than a CPU-only mini PC.
| Situation | Hardware to put first | Guide |
|---|---|---|
| Slow replies on the CPU | A GPU, or a Mac M-series | GPU and eGPU |
| The disk fills up after 3 pulls | A 1–2 TB SSD | NVMe SSD |
| The box is loud at night | A dB reading, plus a fanless profile | Noise and power |
| Several apps (Home Assistant and n8n) | A mini PC with 32 GB of RAM | NAS versus a second mini PC |
A five-week buying path, the careful way
- Week 1. 15 minutes on the living-room PC. Note the peak RAM used, in Task Manager or Activity Monitor.
- Week 2. Open WebUI if several people will use it. Otherwise stay with the Ollama command line.
- Week 3. The hardware selector with your real price per kilowatt-hour (electricity budget).
- Week 4. Read the mini PC comparison or Mac mini M4, depending on the computers you already have.
- Week 5. Order the hardware, plus an SSD if the system disk has less than 500 GB free.
This pace keeps you from buying a 2 TB drive or an RTX before you have confirmed that text is the job. Many households stop at family chat (the private assistant) and never open ComfyUI.
Cooling and where you put the box
A mini PC in a TV stand can overheat. Leave 5–10 cm (about 2–4 inches) of air around the vents. A gaming tower in a house with children needs cables and ports kept out of reach. See GPU and eGPU.
Docker versus a native install, and what it does to the hardware
Docker with Ollama and Open WebUI uses a bit more disk for images, and it makes updates simpler. On a 16 GB mini PC, prefer native Ollama plus Open WebUI, and do not stack other containers. On 32 GB, Docker is comfortable for n8n next to Ollama. See Home Assistant if home automation shares the same machine.
Transcription and images: size the box for the job
Whisper (the software guide and the first steps) mostly uses the CPU, or a CUDA GPU. ComfyUI (FLUX, and the image use) needs VRAM and a large SSD. Do not mix those jobs with a chat-only mini PC. The hardware selector separates a text Budget profile from multimedia Comfort.
Shops and used gear (October 2026 caution)
- New mini PCs and parts: amazon.co.uk, Best Buy, Micro Center, and Newegg. Cross-check the ranges in the table against the price shown when you order.
- Apple: Apple.com and authorized resellers for a Mac mini. See the M4 guide.
- A used GPU (RTX): worth a look if the card was not mining 24 hours a day. Check the remaining warranty and the case power supply (GPU and eGPU).
- LocAIhome prices: no figure is guaranteed. These are October 2026 ranges only. Check them at the store.
Common questions
Can I start without a dedicated graphics card?
Yes, with 32 GB of RAM and a quantized 7B–8B GGUF model. A dedicated graphics card (a GPU) mainly speeds up larger models and image generation.
What is the first concrete step?
List what you will use it for (chat, home automation, transcription, images), then open the LocAIhome hardware selector with your budget and your noise limit.
Mini PC or Mac mini?
An x86 mini PC can take new parts, and an NVIDIA GPU is possible. A Mac mini M4 is quiet and efficient, with unified memory. See the dedicated guide.
Sources
Narrow it down with the hardware selector · French version of this section