How Much RAM Do You Need to Run a Local LLM?

For a small language model used for rewriting and everyday text work, 8 GB of installed RAM is workable and 16 GB is comfortable. For general-purpose local models in the 7 to 8 billion parameter range, 16 GB is the practical floor and 32 GB removes the friction.

The number that actually matters, though, is not what is installed. It is what is free while you are working.

The Rough Arithmetic

A quantized model needs roughly its file size in memory, plus overhead for the context window and the runtime. A reasonable estimate is file size plus about 20 to 50 percent.

Some concrete anchors at four-bit quantization:

A 1.7 billion parameter model is about 1.1 GB on disk and wants roughly 1.7 GB while running. A 4 billion parameter model is about 2.4 GB on disk and roughly 3 GB in memory. A 7 to 8 billion parameter model is around 4.5 GB on disk and 5 to 6 GB in memory. A 13 billion parameter model is around 8 GB on disk and 9 to 10 GB in memory.

Now subtract reality. Windows itself uses 3 to 4 GB. A browser with fifteen tabs happily takes 4 GB. Teams, Slack, an editor, and a PDF viewer add another 2 to 3 GB. On a 16 GB machine you may have 5 GB genuinely free on a normal working afternoon.

That is why a 4 GB model on a 16 GB laptop can still cause your machine to start swapping, and why the smaller model is often the better everyday choice even when the bigger one technically fits.

Check Before You Download

Open Task Manager, go to the Performance tab, and look at Memory. The figure you want is “Available,” not “In use.” Do this while your normal applications are open, not on a freshly rebooted machine, because a fresh reboot flatters you by about 4 GB.

If Available is comfortably above the model’s expected footprint, you are fine. If it is close, you will get generation that starts fast and then stalls as Windows pages memory to disk.

Context Length Costs Memory Too

The other consumer of RAM is the context window, the amount of text the model holds in working memory. This scales with how much you feed it.

For rewriting, this rarely matters. You are handing it an email, not a hundred-page contract. A 4096 token context is ample for the vast majority of work messages and keeps the memory cost small and predictable.

If you plan to summarize long documents locally, budget more, and be aware that long contexts slow generation noticeably as well as consuming memory.

Bigger Is Not Automatically Better

There is a strong instinct to download the largest model that fits. For rewriting specifically, that instinct is usually wrong.

Rewriting is a constrained task. The content already exists; the model is adjusting register, structure, and clarity. Small models handle that well because they are not being asked to reason or recall facts. What you gain from a 13 billion parameter model on a tone rewrite is marginal, and what you lose is several seconds per request and a machine that feels sluggish while it thinks.

The point where model size genuinely matters is open-ended reasoning, code generation, and anything requiring world knowledge. Different job, different hardware budget.

What Wrivio Uses

Wrivio’s Local mode ships two options. The Standard model is roughly 1.1 GB on disk and about 1.7 GB of RAM in use, chosen as the default because it runs acceptably on ordinary office laptops. The larger option is around 2.4 GB on disk and roughly 3 GB in memory for people who want a bit more polish and have the headroom.

Both run in-process on the CPU with no separate server and no GPU requirement. The download is resumable and checksum-verified, and the app shows you the disk and memory figures before you commit to it.

Common Questions

Can I run a local model on 8 GB?

Yes, with a small model and reasonable discipline about what else is open. Under about 2 GB of model footprint is the comfortable zone.

Does the model stay in RAM permanently?

Implementations differ. Wrivio loads the model on first use and caches it, which is why the first rewrite after launch takes longer than the ones after it.

Is a fast SSD a substitute for RAM?

No. Once the system starts paging, generation speed collapses regardless of drive speed.

Does more RAM make it faster?

Only by preventing swapping. Beyond that, memory bandwidth and model size determine speed, not capacity.

Download Wrivio for Windows to run local rewrites with a model that fits comfortably alongside everything else you have open.