Can You Really Run Frontier AI on Your Laptop?

Every time a lab drops the weights for a big model, the same hopeful comment appears: great, now I can run it on my laptop. It is an understandable leap. The download is free, the licence says you can keep it, and the marketing often shows the model running on a desktop. The catch is that free to own and able to run are two very different things, and the gap between them is measured in hundreds of gigabytes of memory.

Where the idea comes from

The confusion has real roots. Small open models genuinely do run on ordinary hardware now. A quantised model with a few billion parameters will sit happily on a decent laptop and answer most everyday questions. Chipmakers have leaned into this, with releases aimed squarely at running capable models on local machines. So the instinct that AI is moving onto personal devices is not wrong. It just does not scale the way people assume.

The hardware math

Take this week's example. Meta's new Muse Spark Open is open-weight and, in Meta's own demo, runs at home. Read the model card, though, and the quantised version wants at least 700GB of memory, while the full-precision build asks for 2TB. Meta's showcase rig is two of Apple's top Mac Studios lashed together, roughly a terabyte of memory, and the configuration it needs does not even ship until later this month.

A typical laptop has 16 or 32GB. A very well-specced one might reach 128GB. You are short by a factor of ten or more, and no amount of patience closes that gap, because the model physically has to fit in memory to run at a usable speed. This is not a software problem waiting for an update. It is arithmetic.

What is actually true

So here is the honest version. You can run a small or mid-sized open model on your own machine today, privately, with no subscription, and that is a genuinely useful thing. You cannot run an 800-billion-parameter frontier model on a laptop, and you will not be able to next year either, unless your laptop quietly grows a terabyte of RAM.

The useful distinction is between open weights and accessible weights. Publishing the weights makes a model free. It does not make it small. The frontier is getting more open, which is worth celebrating, but it is not arriving on your lap any time soon. For that, the cloud is still doing the heavy lifting, and the electricity bill is still landing on someone else's desk.