8 GB is the most common amount of memory in the world, and it is enough for a genuinely useful AI model — as long as you pick the right size and stop expecting it to behave like a data centre.
Your operating system and browser need memory too. On an 8 GB machine, assume about 6.5 GB is realistically available for a model plus its working context. Anything that needs more than that will either refuse to load or push your machine into swapping, which is worse than refusing.
| Model size | File (Q4) | Fits in 8 GB? | Feel |
|---|---|---|---|
| 1–3B | ~1.8 GB | Yes, comfortably | Fast, basic |
| 7–8B | ~4.8 GB | Yes, tight | Slow but usable |
| 13–14B | ~8.4 GB | No | — |
| 30B+ | ~19 GB | No | — |
If you want speed and simple help: a 1–3B model at Q4. Answers arrive quickly and it handles short writing, summarising and tidying text without drama.
If you want actual documents: a 7–8B model at Q4, and accept that a 900-word document will take several minutes. It writes far better than the small ones. Close your browser while it runs.
If you have no graphics card: that is fine. A card makes it faster; it does not make the answers better. The same model file gives the same quality.
Arithmetic. Small models predict numbers rather than computing them, and they are confidently wrong. Ask for the working and check it, or use a calculator.
Recent facts. A model knows nothing after its training cutoff and does not reliably know that. Give it the facts and ask it to work with them.
8 GB runs a real AI assistant for writing and reading, offline and free, today. It does not run a frontier model, and no amount of tuning will change that. Knowing which side of that line your task sits on is the whole skill.
29 pages, plain language, no account needed. What a language model actually is, which kinds exist, what you are really paying for, and how to run one yourself offline.