How Much Disk Space Local AI Really Needs

Local AI is the rare category of software where the download is the product. You are not installing a client that talks to a server, you are installing the model itself, and the model is large.

On a work laptop with a 256GB drive that is already two thirds full, this matters more than the marketing pages admit. Here is what it actually costs.

The Model File Is The Headline Number

A quantized text model for local use typically lands between 600MB and 5GB depending on parameter count and precision.

Wrivio ships two Apache 2.0 Qwen3 tiers. Standard is a 1.7B model at roughly 1.1GB on disk. Best is a 4B Instruct model at roughly 2.4GB. You need one of them, not both, though people who move between mains power and battery often keep both.

Those are the compressed, four-bit numbers. The same models at higher precision would be several times larger for a quality difference you would struggle to notice in a rewritten paragraph, which is why nearly every local tool ships quantized weights by default. Quantization explained for writers covers why that trade is usually right.

Downloads Need Roughly Double The Space, Briefly

This is the part that catches people out on a nearly full drive.

A resumable download writes to a temporary file and only becomes the final model after its checksum is verified. During that window you are holding the partial file, and depending on how the tool is implemented you may briefly hold both the partial and the final copy.

Plan for the model size plus the same again as working room. A 2.4GB model wants roughly 5GB free to install cleanly, even though it settles at 2.4GB.

If a download fails repeatedly near the end on a full disk, this is almost always why, and the error message rarely says so.

Where It Lives Determines Whether IT Can See It

On Windows, application data of this kind usually sits under your user profile in AppData, not in Program Files. That has two consequences worth knowing.

First, it counts against your profile, which on a managed machine may be synced, backed up, or quota’d. A roaming profile that suddenly grows by 2.4GB will get noticed.

Second, uninstalling the application does not always remove it. Model files are user data by most installers’ reckoning, so they can survive an uninstall and sit there indefinitely. If you have trialled several local AI tools, check for orphaned model files before concluding your disk is mysteriously full.

Store-packaged applications on Windows historically complicated this further, because package data could live in a virtualised location that behaved unexpectedly on uninstall or reset. Wrivio moved its data out of that virtualised path specifically so a Store install and a direct download share one predictable location and neither loses your model on an app reset.

Context Length Costs Memory, Not Disk

A common confusion is worth clearing up, because it changes what you should buy.

The model file is fixed on disk. The context window, meaning how much text the model can consider at once, costs RAM at runtime and grows with the amount of text you feed it. A tool advertising a very long context is making a memory claim, not a storage one.

For rewriting work this is rarely a constraint, because you are pasting paragraphs rather than books. It matters if you plan to summarise long documents locally. How much RAM to run a local LLM is the relevant piece there.

Reclaiming Space Without Breaking Anything

If you need the space back, the order of operations is straightforward.

Delete unused model tiers first. If you downloaded a 4B model, decided the 1.7B was fine, and never went back, the larger file is pure cost. Most tools expose this in settings; Wrivio lists both tiers with a delete option next to each.

Remove orphaned files from tools you no longer use. This is usually the biggest single win, because those files are invisible in the sense that no running application references them.

Do not delete the model of the tool you are actively using and expect it to degrade gracefully. Most implementations gate the feature behind the model being present and will simply offer to download it again, which is the correct behaviour but not what you wanted at 9am.

And consider whether you need local at all for everything. A hybrid local and cloud workflow keeps one small model on disk for confidential work and uses the cloud for the rest, which is often the right answer on a constrained machine.

Common Questions

How much disk space does a local AI writing model need?

Between roughly 1GB and 2.5GB for the model tiers most tools ship, plus the same again as temporary working room during the download. Budget around 5GB free to install a 2.4GB model comfortably.

Does uninstalling the app remove the model?

Often not. Model files are treated as user data by many installers and can survive an uninstall, so check the application data directory if you have trialled several local tools and your disk is unexpectedly full.

Does a longer context window need more disk space?

No, it needs more RAM. The model file size is fixed; context length is a runtime memory cost that grows with how much text you feed the model at once.

Can I move the model to an external drive?

Sometimes, but it is rarely a good idea. Inference reads the model repeatedly and external drives add latency to exactly the path that decides how fast generation feels.

Download Wrivio for Windows and start with the Standard tier at roughly 1.1GB, which is enough for everyday rewriting and leaves your disk alone.