Keeping it tidy

Model files pile up quietly

A model is a file, and it is a big one. You download three to compare them, keep the one you liked, and forget the other two. Do that for a few months and a lot of your disk is gone.

Ollama's own installation notes put the scale plainly. Once the app is installed, they say, you need extra room for the models, which can be tens to hundreds of gigabytes.

Nothing warns you. The files sit in a folder your tool owns, out of the way, and they do not shrink.

Two commands do the tidying. Run ollama ls to see what you have. Run ollama rm with a model name to delete one. Deleting a model is safe. If you want it back, you download it again.

The next two sections are the folders those files are in, and the four commands worth knowing by heart.

Backed by Ollama CLI Reference: ls, ps, rm and stop, Ollama on macOS: filesystem requirements and file locations, Ollama on Windows: filesystem requirements and file locations.

The actual folders

Every tool has one place it keeps things. Knowing it turns a mystery into a folder you can look at.

Ollama uses a hidden .ollama folder. On a Mac that is in your home folder, and it holds both the models and the configuration. Its logs sit beside them, in a logs folder, as app.log and server.log.

On Windows the same .ollama folder is under your user folder, and the logs are somewhere else, under the Ollama folder in your local application data. On Linux the models are under the shared ollama folder in /usr/share.

LM Studio keeps its models in a .lmstudio folder in your home folder, and it lays them out the way Hugging Face does. A folder for the publisher, a folder for the model, and the file inside.

Their own example shows you exactly what to expect. A folder called infra-ai. Inside it, a folder called ocelot-v1. Inside that, a file called ocelot-v1-instruct-q4_0.gguf.

Read that file name from the end. The q4_0 says how hard it was squeezed, since LM Studio use the letter Q for quantization on their download page too. The instruct says it is a chat version rather than a base one.

Backed by Ollama FAQ: where models are stored, and how to upgrade, Ollama on macOS: filesystem requirements and file locations, Ollama on Windows: filesystem requirements and file locations, LM Studio: the expected models directory structure, LM Studio: Download an LLM, and changing the models directory.

Four commands, and the two people confuse

There are two questions you will ask over and over. What have I got, and what is running now.

In Ollama, ollama ls answers the first and ollama ps answers the second. LM Studio has the same pair at the terminal: lms ls lists the models on disk, and lms ps lists the ones loaded into memory.

The other two commands both free something up, and they free different things.

ollama rm deletes the file. That is disk space back, and the model is gone until you download it again.

ollama stop unloads the model from memory. The file stays exactly where it was. That is memory back, and the model is still yours.

People reach for rm when they meant stop, usually while trying to make a slow session faster. Slow sessions are a memory problem, so stop is almost always the one you wanted.

Backed by Ollama CLI Reference: ls, ps, rm and stop, lms, LM Studio's command line tool: ls, ps and runtime.

Two things update, not one

People get caught by this. Updating your local AI app does not make your model newer.

The app is the program that reads the model file. It gets fixed and improved often. Ollama on Mac and Windows downloads its own updates and asks you to restart. On Linux you run the install script again.

The model is the file. It does not change. Meta's own model card for Llama 3.1 calls it a static model trained on an offline dataset. When a better version arrives, it arrives as a different file with a different name.

So an update to the app and an update to a model are two separate jobs. If you want a newer model, you have to go and fetch it.

Nothing about a model file wears out either. A model you downloaded a year ago answers exactly as it did on the day you got it, for better and for worse.

Backed by Ollama FAQ: where models are stored, and how to upgrade, Llama 3.1 model card, lms, LM Studio's command line tool: ls, ps and runtime.

Know which model you are talking to

This matters more than it sounds. If you do not know what produced a result, you cannot repeat it.

Model names carry facts. They usually tell you the family, the size, and how hard the file was squeezed. Two names that differ by a few letters can be very different files.

Many apps let you switch models in the middle of a session, or pick one for you. So the answer you liked last week may not have come from the model that is loaded now.

Run ollama ps to see what is running at this moment. Their own example line reads gemma4:latest, then an id, then 9.6 GB, then 100% GPU, then 131072.

Every one of those is worth reading. The size is what it is costing you. The 100% GPU is your speed check from the last lesson. The last number is the window it was given.

Backed by Ollama CLI Reference: ls, ps, rm and stop, Ollama: Context length, with a sample ollama ps output.

A ten minute tidy

Once a month is enough. Put it next to some other dull task you already do.

• List what you have downloaded, and read the sizes.
• Delete anything you have not used since the last tidy.
• Update the app, and note the version somewhere you will find it.
• Write down the exact name of the model you rely on for real work.
• Ask that model one question you know the answer to, so a bad change does not go unnoticed.
• Glance at the logs folder. If something has been failing quietly, that is where it said so.

Backed by Ollama on macOS: filesystem requirements and file locations.