How much RAM do I actually need in 2026?

16GB is the floor for a new machine, 24GB to 32GB suits development, design and occasional video work, and 32GB to 64GB is for heavy video, 3D, virtualization or running large models locally. The old advice that 8GB is fine for ordinary use no longer holds, because the baseline stack of an OS, a browser, and a video call now consumes most of an 8GB machine on its own.

Is 8GB of RAM still enough?

Only for genuinely light use on a cheap or secondary machine, such as a Chromebook or a browsing-and-email laptop. On a primary computer the OS plus a dozen tabs plus a call already runs 7GB to 9GB, so an 8GB machine spends its life compressing and swapping, which shows up as a system that feels sticky rather than one that visibly fails.

Can I upgrade my laptop RAM later?

Usually not, and you have to check the specific model before you buy. Apple Silicon unified memory sits on the processor package and is never upgradeable, most thin Windows laptops solder LPDDR5X for power and space reasons, and only machines that list SODIMM slots or the newer CAMM2 standard can be changed after purchase.

How much RAM do I need to run AI models locally?

Budget roughly the size of the model's quantized weights plus a few GB of overhead, on top of your normal working set. At 4-bit quantization that is about 5GB for an 8B model, 9GB for 14B, 20GB for 32B and 40GB for 70B, which in practice means a 16GB Mac handles the 8B class comfortably and a 32GB machine handles the 32B class.

Why does my computer get slow over time?

Slowdowns usually happen when the working set of open apps stops fitting in RAM and the OS starts compressing memory and swapping to disk. Disk is far slower than RAM, and that paging, rather than the CPU or the drive itself, is what people experience as a computer getting slow.

How do I check whether I have enough RAM?

On macOS open Activity Monitor, go to the Memory tab, and watch Memory Pressure during real work: green means you have enough, while yellow or red means the system is compressing and swapping to cope. On Windows use Task Manager's Performance tab and watch the memory graph across a normal working day rather than at idle.

How Much RAM Do I Actually Need? (2026)

For years the contrarian answer to this question was “8GB is fine, you'll never use the rest, save the money.” That answer was right for a long time and it is not right anymore, which is worth saying plainly because a lot of pages ranking for this query, including an earlier version of this one, still give it.

Two things changed. The baseline software stack got heavier, and the memory in most machines stopped being something you can add later. The second one matters more than the first, and it is the question to settle before you pick a number at all.

What RAM Is Actually For

RAM is the working memory of a computer. When you open Chrome, the browser doesn't run from disk; it gets loaded into RAM where the CPU can read and write to it at memory-speed rather than disk-speed. The OS, every running app, every open browser tab, and every piece of working data lives in RAM. When you run out of RAM, the OS first compresses what it can in memory, then starts swapping pages out to disk, which is orders of magnitude slower than RAM even on a fast NVMe SSD.

That swap-to-disk moment is what people experience as their computer "getting slow" over time. It's not the CPU. It's not the disk. It's the moment the working set of open apps stops fitting in RAM and the OS starts paging.

Start Here: You Probably Cannot Add It Later

Every other consideration is downstream of this one, and it is the reason the advice that was right a few years ago is wrong now.

Thin laptops overwhelmingly solder their memory to the board. Apple Silicon goes further: memory sits on the same package as the processor as unified memory, and it is not upgradeable by Apple, by a repair shop, or by anyone. A large share of Windows ultrabooks use soldered LPDDR5X for the same reasons Apple does, namely power and board space. Some bigger or more serviceable machines still ship replaceable SODIMM slots, and the newer CAMM2 standard is starting to bring upgradeable memory back to thin designs, but you have to check the specific model rather than assume.

So the question is not “how much RAM do I need,” it is “how much will I need before I replace this machine.” If the RAM is soldered, being wrong costs a laptop. If it is socketed, being wrong costs a memory stick and twenty minutes. Find out which you are buying before you decide the number, because the right answer differs by a whole tier.

Why this matters

Check the listing for “soldered,“ “onboard,” “LPDDR5” or “unified memory,” all of which mean not upgradeable. “SODIMM,” “CAMM2,” or a stated number of memory slots means you can change it later. On a soldered machine, buy one tier above what you think you need. On a socketed one, buy what you need today and upgrade if you are wrong.

What Apps Actually Use

Chrome with a dozen tabs runs roughly 2GB to 4GB, and more if any of those tabs are applications rather than documents, which for Docs, Figma, or a webmail client they are. Windows 11 wants 3GB to 4GB for itself; current macOS is similar. Add a video call, a chat client, a music app, and a password manager, and a genuinely ordinary desk setup sits at 7GB to 9GB before you open the thing you actually do your job in.

That is the number that moved. It is why the old “8GB is fine for most people” advice, which was true when the same stack cost 5GB, no longer is. On an 8GB machine today, the baseline load alone fills the memory and the system compresses and swaps continuously to keep up. Modern memory compression is good enough that this is often not catastrophic, which is exactly what makes it insidious. It shows up as a machine that feels sticky rather than one that visibly fails.

Apple made the same call with its own hardware. The MacBook Air moved to 16GB as the base configuration with the M4 generation, after years of shipping 8GB, and the M5 kept it there.^[1] When the vendor with the strongest memory-efficiency story and the most to gain from upselling raises its own floor, that is the clearest signal available on where the baseline actually sits.

The Recommendation, By What You Do

Browsing, email, Office, streaming: 16GB.Not 8GB, not anymore. 16GB is the floor for a machine you want to still feel fine in three years, and the upcharge from 8GB is usually modest on a Windows laptop. The exception worth naming honestly: if you are buying a cheap secondary machine or a Chromebook for genuinely light use, 8GB still works, and pretending otherwise wastes people's money.

Software development, design, occasional video: 24GB to 32GB. 16GB is workable and was the standard recommendation for years, but it is now the tier where you notice the ceiling: an IDE with language servers, a browser, a container stack, and a call all at once will find the edge of 16GB. If the machine is soldered, this is where the extra tier earns its money.

Heavy video, 3D, virtualization, local AI models: 32GB to 64GB. Below 32GB you will hit swap under normal load. Virtual machines and large in-memory datasets are the classic drivers here, and local models are the new one.

The New Variable: Running Models Locally

This is the part no RAM guide written before 2024 accounts for, and it is now the most common reason an otherwise adequate machine runs out of memory.

A local language model has to fit in memory to run at a usable speed, and the rough figure is the size of its quantized weights plus a few GB of overhead for context. At the common 4-bit quantization, an 8B model needs somewhere around 5GB, a 14B model around 9GB, a 32B model around 20GB, and a 70B model around 40GB. Those sit on top of your normal working set, not instead of it.

On Apple Silicon this interacts with unified memory in a way that surprises people. The GPU and CPU draw from one shared pool, so a model loaded onto the GPU is consuming the same RAM your browser wants, and macOS limits how much of the total a GPU process may claim. Practically: a 16GB Mac comfortably runs models in the 8B class and struggles above that, a 32GB Mac handles the 32B class, and the larger configurations are what people buy specifically to run bigger models at home. If that is on your roadmap at all, size for it at purchase, because on these machines there is no later. There is more on what that stack looks like in running AI agents on-device.

Check What You Actually Use First

Before paying for a tier, measure. On Windows, open Task Manager, go to the Performance tab, and look at the memory graph during a normal working day rather than at idle. On macOS, open Activity Monitor, go to the Memory tab, and watch Memory Pressure, which is the number that matters. Green means you have enough. Yellow and red mean the system is compressing and swapping to cope, and that is your answer regardless of what the raw GB figure says.

macOS in particular will happily report almost all memory as “used” on a perfectly healthy machine, because unused RAM is wasted RAM and the OS fills it with cache it can drop instantly. People see that number, panic, and buy a tier they did not need. Memory pressure is the signal; the used figure is not.

Takeaway

In 2026 the floor is 16GB, not 8GB, and Apple raising its own base configuration to 16GB is the tell. But the number matters less than whether the memory is soldered: on a machine you can never upgrade, buy one tier above your estimate, and on one with real memory slots, buy what you need now. If you intend to run models locally, size for the model at purchase, because that requirement lands entirely on top of everything else.

Sources and further reading

1.ReportingHow much RAM does your PC actually need in 2026?. Source for the current 16GB baseline and Apple's move to 16GB as the MacBook Air base configuration.