You do not need a graphics card, a subscription, or an account. A normal laptop can run a useful AI model today. Here is what it takes, and what it will not do.
Install one program, download one model file, load it, and ask it something. Turn your Wi-Fi off and it keeps answering. That is the whole of it — about twenty minutes, no account.
Almost everyone assumes the processor matters most. It does not. Running a model means reading the whole model file to produce each word, so the constraint is how fast memory can feed the processor — and how much memory you have.
| Your machine | What it means in practice |
|---|---|
| 8 GB of memory | Small models run fine. Larger ones will not fit alongside your browser. |
| 16 GB | Comfortable for everyday models. This is the sweet spot for ordinary use. |
| 32 GB and up | Bigger models become possible, not fast. |
| A graphics card | Much faster. Not required, and the model quality is identical. |
One detail that surprises people: two memory sticks instead of one can roughly double the speed, because two sticks let the processor read in parallel. Many laptops ship with a single stick, which is why an identical-sounding machine can feel twice as fast.
Models are named by size in billions of parameters. For a normal laptop:
You will also see Q4 in file names. That is quantisation: the model compressed to about four bits per parameter. A Q4 file is roughly 0.6 GB per billion parameters, so an 8B model is about 4.8 GB on disk. Q8 is closer to the original and roughly twice the size.
On a processor-only laptop, expect something like this — and treat it as a range, because real machines vary by 30–50% with background programs and cooling:
| Model | Single memory stick | Two sticks |
|---|---|---|
| 3B, Q4 | ~7 words/second | ~14 words/second |
| 8B, Q4 | ~3 words/second | ~6 words/second |
Three words a second is slow. It is also completely private, free forever, and works on a train. Whether that is a good trade depends entirely on what you are doing.
It will not match the largest cloud models on hard reasoning. It cannot know anything after its training cutoff. It still invents facts with total confidence — that is a property of how these models work, not a bug in the small ones.
And it needs your computer to stay awake, so a long task on a laptop means plugging it in.
Not installation — choosing. There are hundreds of model files and no guidance about which one suits your machine, so people download something too large, watch it crawl, and conclude local AI is a toy. Pick by your memory figure first, then by what you want it to do.
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.