Which computer to buy for local AI in 2026
For local AI the most important part is the graphics card, above all the amount of its memory (VRAM). The model has to fit into memory, otherwise it runs much more slowly. Smaller models need a card with 16 GB of VRAM, mid-sized models 24 to 32 GB, and the largest open models several cards or a workstation with large unified memory.
Why the graphics card
Language models perform an enormous number of matrix calculations. Graphics cards handle them in parallel and many times faster than a CPU. Response speed depends most on memory bandwidth, which is why VRAM matters more than raw chip performance.
How much memory a model needs
Model size is given in billions of parameters (e.g. 8B, 32B, 70B). For local use, models are usually quantised, i.e. shrunk with minimal loss of quality. As a rough guide with common 4-bit quantisation:
| Model size | VRAM needed (approx.) | Good for |
|---|---|---|
| 7–9 billion (7–9B) | 6–8 GB | Simple queries, summaries, sorting e-mails |
| 12–14 billion | 10–12 GB | Good multilingual output, writing, working with documents |
| 27–32 billion | 18–24 GB | More demanding analysis, agents, programming |
| 70 billion and more | 40 GB and more | Quality close to cloud models |
Add headroom for the context, i.e. the length of text the model works with at once. Long documents and several simultaneous users need more memory.
Three example builds
1. Office AI workstation
A graphics card with 16 GB of VRAM, 32–64 GB of RAM and a fast NVMe drive. Handles models up to about 14 billion parameters. Ideal for an individual or a small team: writing, summarising, answering from documents.
2. Company AI server
A card with 24–32 GB of VRAM (or two cards), 64–128 GB of RAM and an NVMe drive of 2 TB or more. Models up to 32 billion parameters with long context, several users at once, AI agents over company data.
3. High-end workstation for large models
Multiple graphics cards or professional cards with large memory, or a computer with unified memory where CPU and GPU share tens to hundreds of GB. For 70B+ models and heavy workloads.
Things not to forget
- Power supply and cooling: powerful cards draw a lot of power, the PSU needs headroom and the case good airflow.
- Noise: if the workstation sits in an office, quieter cooling is worth it.
- Storage: models range from a few to tens of GB, and a fast NVMe drive shortens loading.
- Software: tools such as Ollama, llama.cpp or vLLM run the models, with a web interface for users such as Open WebUI.
Is it worth it?
Local AI makes sense when you work with sensitive data, pay AI subscriptions for a larger team or need operation without limits and without the internet. For occasional use the cloud is usually cheaper. Often a combination of both is best.
If you want a build designed for you, including installation of models and AI agents, see our Local AI & AI workstations service.