How to delete Ollama models and free up space on a Mac

Ollama keeps every model you download in one folder, ~/.ollama/models, and a model stays there until you remove it. To delete Ollama models and get the space back, remove them through Ollama itself. It knows which files each model uses, and it deletes only the ones no other model still needs.

The catch is that Ollama's Mac app has no button for this, as of version 0.40. Removing a single model happens through Ollama's command-line tool, which comes with the app, or through its local API. In Finder, only two moves are safe: putting the whole models folder in the Trash, or moving it to another disk.

This post explains how the folder is built, how large models really are, how to see what you have, what removing a model gives back, and what else Ollama keeps on your Mac. For LM Studio, Hugging Face and the other tools, see where local AI tools keep their models.

Where Ollama keeps models on a Mac

Ollama keeps its models and configuration in a hidden folder in your home folder, ~/.ollama. The models are in the models folder inside it. The app in your Applications folder holds only the program, so updating or reinstalling Ollama leaves your models where they are.

To open the folder, choose Go › Go to Folder in Finder, type ~/.ollama/models and press Return. The tilde stands for your home folder. Folders whose names start with a period are hidden in Finder, but Go to Folder opens them anyway.

What's inside the models folder

You'll find two folders, and in recent versions a third:

• blobs holds the data: each model's weights, plus small files for its settings, prompt template and license. Almost all the space is here.
• manifests holds one small file for each model name and tag you have, such as manifests/registry.ollama.ai/library/llama3.1/8b. A manifest lists the blobs a model is made of.
• manifests-v2, added in Ollama 0.40, holds the same kind of entries under manifests-v2/ollama.com/library. Ollama may keep entries in the older folder too, so an earlier version can still find your models.

If you changed Model location in Ollama's settings, the same three folders are wherever you pointed it, and the folder in your home folder may be empty or missing.

Why the file names are long hashes

Each file in blobs is named sha256- followed by 64 letters and digits. That string is the file's SHA-256 digest: a fingerprint worked out from its contents. Ollama checks every download against it, and throws away a file that doesn't match so it can be downloaded again.

Naming files by their contents has a useful side effect. Two models that contain an identical file both point to the same blob, so it's stored once. It also means the names tell you nothing about which model a file belongs to. Only the manifests say that.

How models share layers

Ollama calls the blobs a model is made of its layers. Models share layers more often than you'd expect:

• Several names for one model. In Ollama's library, llama3.1, llama3.1:8b and llama3.1:8b-instruct-q4_K_M are one download with one ID. Pull all three and the weights are stored once.
• Models you create. A model built from a Modelfile that starts with FROM llama3.1 reuses Llama 3.1's weights and adds a few small layers of its own, such as a system prompt.
• Copies. Copying a model under a new name, as Ollama's guide to sharing models does, adds a manifest and nothing else.

This sharing is why removing a model can free much less than its size, and why removing files by hand is risky. Both come up again below.

Path	What it holds	How big
~/.ollama/models/blobs	Model weights and their small parts, named by hash	Gigabytes per model
~/.ollama/models/manifests	One small file per model name and tag	Tiny
~/.ollama/models/manifests-v2	The same, in Ollama 0.40 and later	Tiny
~/.ollama/logs	The app's and the server's logs	A few files
~/.ollama/id_ed25519	The key that links this Mac to an ollama.com account	Under a kilobyte
/Applications/Ollama.app	The app, its server and its command-line tool	About 200 MB to download

How big Ollama models are

A model's size depends mostly on two things: how many parameters it has, and how precisely each one is stored. The number in a tag, such as 8b or 70b, is the parameter count in billions. Ollama's library shows every tag's size before you pull it, so you can check before you download.

• nomic-embed-text0.3 GB
• llama3.2 (3B)2 GB
• gemma3 (4B)3.4 GB
• llama3.1 (8B)4.9 GB
• qwen3 (8B)5.2 GB
• gemma3:27b17 GB
• deepseek-r1:32b20 GB
• llama3.3 (70B)43 GB
• gpt-oss:120b65 GB

Sizes as Ollama's library lists them in October 2026, for the default build of each tag. Larger quantizations of the same models take more.

Sizes add up quickly. Gemma 3 27B, DeepSeek-R1 32B and Llama 3.3 70B together come to 80 GB. The arithmetic behind sizes like these is in how much storage local models need.

What quantization does to the size

Most models in Ollama's library are quantized: each weight is stored in fewer bits than the original, which makes the file smaller at some cost in quality. The default tag is usually a 4-bit build, marked q4_K_M. Other builds are offered under longer tags, and each one is a separate download.

Here's what that looks like for one model, Llama 3.1 with 8 billion parameters:

Tag ending	What it is	Size
q2_K	The smallest build, with the most loss	3.2 GB
q3_K_M	A 3-bit build	4.0 GB
q4_K_M	A 4-bit build, and what the default tag points to	4.9 GB
q5_K_M	A 5-bit build	5.7 GB
q6_K	A 6-bit build	6.6 GB
q8_0	An 8-bit build, close to the original	8.5 GB
fp16	The original 16-bit weights, not quantized	16 GB

The unquantized build is more than three times the default. If you pulled a few builds to compare them, keep the one you use: the others are full copies, not small variations, and they share no weights.

Why some models show a range of sizes

Some newer models list a range instead of one number. Gemma 4's default tag, for example, shows 6.6 GB to 9.5 GB. The library holds more than one build of these models, for Ollama's different engines, and a download fetches only the one that suits your Mac.

On a Mac with Apple silicon, recent versions of Ollama prefer the MLX build when there is one. For Gemma 4's default tag, the MLX build is the 9.5 GB one. On an Intel Mac, you'd get a different build.

Do cloud models take space?

Names that end in cloud, such as gemma4:cloud, run on Ollama's servers. Ollama's documentation says they don't need to be downloaded, so they take almost nothing on your Mac. They do need you to sign in, which uses the key described later in this post.

How to see which models you have and what they take

There's no single screen that shows each model with its size, unless you use Ollama's command-line tool. Each of the other places shows part of the picture.

In Ollama's app

In a chat, click the model's name in the message box to open the model menu. Models already on your Mac appear without a download arrow. A model with the arrow is downloaded when you first use it, and a cloud icon marks models that run on Ollama's servers.

The menu doesn't show sizes, and it has no way to remove a model. Its search field, Find model, helps when the list is long.

In Finder

Open ~/.ollama/models/manifests, or manifests-v2, with Go › Go to Folder, then open the folders inside until you reach the model names. Each model is a folder, and each tag you have is a file inside it. That tells you which models you have, but not how big each one is.

For the total, select the models folder and choose File › Get Info. Finder counts each blob once, so this is the real space your models take. To see the largest files, open blobs, choose View › as List and click the Size column. The largest blobs are model weights, though their names won't tell you whose.

With Ollama's command-line tool

Ollama's list command prints each model's name, ID, size and when it last changed. It's the only place that puts a model's name next to its size. Two names with the same ID are the same model, stored once.

In Storage settings

Choose Apple menu › System Settings, click General, then Storage. Ollama the app is counted under Applications. Ollama's models usually land in System Data, the category for everything macOS can't place elsewhere, which is one reason System Data can grow so large.

Why the numbers don't add up

The size Ollama lists for a model counts every layer it uses, even layers another model shares. Add up the list and you count shared layers more than once. Finder's Get Info counts each file once, so it's usually lower.

Here's a worked example of a Mac with six model names:

Model	Size Ollama lists	Shares its weights with
llama3.1:latest	4.9 GB	llama3.1:8b and my-assistant
llama3.1:8b	4.9 GB	llama3.1:latest and my-assistant
my-assistant, made from Llama 3.1	4.9 GB	Both Llama 3.1 names
gemma3:27b	17 GB	Nothing
deepseek-r1:32b	20 GB	Nothing
nomic-embed-text	0.3 GB	Nothing

The list adds up to about 52 GB. On disk, the three Llama 3.1 names share one set of weights, so the folder holds about 42 GB. That 10 GB gap is space no removal will ever give back, because it was never used.

How to delete Ollama models

Ollama decides what to delete, because only it reads the manifests. Your choice is mostly which route to use, and how much you want to remove.

What Ollama's app offers

Nothing for removing a model. Ollama's app, as of version 0.40, has a chat window, a model menu and a Settings window, and none of them can delete a model. Its Settings window can change where models are stored, covered below.

Deleting a chat in the app's sidebar removes the conversation, not the model it used.

Ollama's command-line tool

The first time the app starts, it asks to add a link to its command-line tool, so the tool works in Terminal. That tool's remove command takes a model's name, as the list command shows it, and removes it. Ollama's CLI reference documents it. There's no window for it: it's typed in Terminal.

Other apps that use Ollama, such as chat front ends, talk to it through its local API, which has the same delete call. If one of them can remove models, it's asking Ollama to do it, with the same result.

Removing every model at once, in Finder

If you want all your models gone, Finder can do it without breaking anything, because no model is left with missing pieces:

1. Quit Ollama: click its icon in the menu bar and choose Quit Ollama.
2. Make sure you have the Modelfiles, and any GGUF files, of models you made yourself. They're wherever you saved them, not in this folder.
3. In Finder, choose Go › Go to Folder, type ~/.ollama and press Return.
4. Drag the models folder to the Trash. Leave the .ollama folder itself: it also holds your key and logs.
5. Open Ollama again. It makes a new, empty models folder by itself.
6. Empty the Trash when you're sure. Until then, the space isn't freed, and you can still put the folder back.

Can you delete files in the blobs folder?

No. You can't tell which model a blob belongs to, and a blob may belong to several. Remove one and every model that uses it stops loading, while its manifest still says it's there. Ollama can't use that model until you download it again.

Removing a model's manifest by hand isn't a documented way to remove it either. Leave the inside of the models folder to Ollama.

Way	What it removes	When the space comes back	Can you undo it?
Ollama's remove command	One model name, and the layers nothing else uses	Right away	No. Download it again
The models folder in the Trash, with Ollama quit	Every model	When you empty the Trash	Yes, until you empty the Trash
Single blobs in Finder	Parts of models you can't identify	Partly	It breaks the models that used them
Ollama in the Trash	Only the app	When you empty the Trash	Yes, until you empty the Trash
Deleting a chat in Ollama's app	The conversation	Almost nothing	No

What removing a model does

Removing a model through Ollama is quick, because it deletes files rather than moving them. Here's what happens, in order, and what it means for you.

What's deleted, and what stays

If the model is loaded in memory, Ollama's tool unloads it first. Ollama then deletes the model's manifest, and checks each of the model's layers against every other manifest. A layer no other model uses is deleted. A layer another model still uses stays.

Your other models, your chats, your settings and your key are untouched.

Why you may get back less than the listed size

Because shared layers stay. In the worked example above, removing llama3.1:latest frees nothing at all: llama3.1:8b is the same model. Removing both Llama 3.1 names still frees almost nothing, because my-assistant uses the same weights. Only removing all three gives back the 4.9 GB.

When a removal frees less than you expected, look for other names with the same ID in the list, and for models you created from the one you removed.

Is there an undo?

No. Ollama deletes the files directly, without the Trash, and it doesn't ask for confirmation. A model from Ollama's library comes back the next time you pull it, or the next time you choose it in the app's model menu, but that's the whole download again.

On a slow connection, that matters for the large ones. A 43 GB model can take hours to download again.

Models you created yourself

A model you made from a Modelfile can't be downloaded again, unless you shared it on ollama.com. To get it back, create it again from the same Modelfile, which needs whatever its FROM line points to: a model from the library, which downloads again, or a GGUF or Safetensors file, which you must still have.

Before removing a model you made, check that you have its Modelfile and the files it was built from.

What happens to your chats

Ollama's app keeps your conversations in its own database, separate from the models. Removing a model leaves those chats in the sidebar. Choose that model again to continue one, and the app downloads it first.

Space Ollama gives back by itself

Ollama tidies its models folder in a few situations, which is why the folder doesn't usually fill with forgotten pieces.

Updating a model

Pulling a model you already have checks for a newer version. Ollama downloads only the layers that changed, then removes the old layers nothing uses any more. So updating a model doesn't keep the old version around under the same name.

A different tag is a different model, though. Pulling llama3.1:70b after llama3.1:8b adds a second model; it doesn't replace the first.

Unfinished downloads

A download you stop, or one that fails when the disk fills up, leaves partial files in blobs, with partial in their names. Pulling the same model again picks up where it left off.

If you don't, Ollama removes them later. Each time its server starts, it deletes partial files and layers no model uses, once they're more than an hour old. Quitting Ollama and opening it again starts the server afresh.

When the tidy-up doesn't run

Two things stop it. One is an environment variable, OLLAMA_NOPRUNE, that turns it off; it has no effect unless someone set it. The other is a manifest Ollama can't read: it then skips the tidy-up altogether and notes "corrupt manifests detected" in its server log.

If the models folder stays far larger than your models, look for that line in ~/.ollama/logs/server.log. The log suggests pulling the affected model again, or removing it.

Stopping a model frees memory, not disk

After you use a model, Ollama keeps it in memory for five minutes by default, so the next answer starts faster. Stopping a model, or waiting, frees that memory. It doesn't change anything on the disk: the model stays downloaded until you remove it.

The reverse is also true. Removing models doesn't make your Mac faster. It gives you disk space back, and that's all.

Moving the models folder to another disk

If you'd rather keep your models than remove them, Ollama's app can keep them somewhere else, such as an external drive. Ollama's macOS documentation suggests this when your home folder lacks the space.

Ollama's Model location setting

Click Ollama's icon in the menu bar and choose Settings, or choose Ollama › Settings while its window is open. Under Model location, click Browse and choose a folder. Ollama restarts its server and uses that folder from then on: new downloads go there, and the model menu lists what's in it.

Your existing models don't move

Changing the setting doesn't copy anything. Models in the old folder stay there, still taking space, and Ollama stops listing them. To bring them along:

1. Quit Ollama from its menu bar icon.
2. In Finder, copy the whole models folder, blobs and manifests together, to the other disk. Never copy only blobs.
3. Open Ollama, open Settings, and set Model location to the copied folder.
4. Check that your models appear in the model menu without download arrows.
5. Move the old ~/.ollama/models folder to the Trash, and empty it when you're sure.

When the other disk isn't connected

If the folder isn't there when Ollama starts, for example because the drive isn't plugged in, Ollama quietly uses ~/.ollama/models on your startup disk instead. Anything you download then lands on the startup disk, and models on the other drive don't appear.

Connect the drive before opening Ollama, and if it starts without it, quit Ollama and open it again once the drive is connected.

If you run Ollama without the app

Homebrew offers Ollama two ways: as the same app, and as a formula with only the server and command-line tool. Without the app there's no Settings window. The location then comes from an environment variable, OLLAMA_MODELS, which Ollama's FAQ documents; there's no interface for it.

Ollama's other files

Models are almost all of Ollama's space, but not all of its files. Most of the rest is small, and some of it matters more than its size.

Logs

~/.ollama/logs holds app.log from the app and server.log from the server that runs models. Each starts fresh when the app or the server starts, and Ollama keeps up to five older copies of each, so old logs don't pile up.

They help only when something goes wrong, and Ollama's troubleshooting guide points to them first. You can move them to the Trash while Ollama is quit, and it writes new ones.

Downloaded updates

Ollama downloads updates in the background, into ~/Library/Caches/ollama. An update waits there until you choose Restart to update in Ollama's menu bar menu, or until Ollama installs it by itself when it starts at login. Ollama removes the download once it's installed, and turning off Auto-download updates in Settings stops it downloading them.

Chat history

The app's chats, its settings and the images you attach to chats live in ~/Library/Application Support/Ollama. That's your own data, not a cache. Delete chats you don't need inside the app, and keep the folder unless you're uninstalling Ollama.

The key in the .ollama folder

id_ed25519 and id_ed25519.pub are a key pair Ollama creates when its server first starts. The public half links this Mac to your ollama.com account, for signing in, running cloud models and sharing models you make. If you remove the pair, Ollama makes a new one, and you sign in again.

Path	What it is	Before you remove it
~/.ollama/logs	The app's and the server's logs	Quit Ollama first; it writes new ones
~/.ollama/history	Prompts you typed in the command-line tool	Only that history is lost
~/Library/Caches/ollama	An update Ollama downloaded and hasn't installed	It downloads the update again
~/Library/Application Support/Ollama	Your chats, their images and the app's settings	Your own data; keep it
~/.ollama/id_ed25519	The key that links this Mac to ollama.com	A new one is made; sign in again
~/Library/WebKit/com.electron.ollama	Data for the app's window	Only when uninstalling

Uninstalling Ollama completely

Dragging Ollama to the Trash removes the app and nothing else. Your models stay in ~/.ollama, sometimes taking tens of gigabytes long after the app is gone. Ollama's own macOS documentation lists everything to remove for a complete uninstall.

Removing it step by step

1. Quit Ollama: click its icon in the menu bar and choose Quit Ollama.
2. In Finder, open Applications and drag Ollama to the Trash. macOS may ask for an administrator's name and password.
3. Choose Go › Go to Folder, type ~/.ollama and press Return. Choose Edit › Select All and move everything to the Trash: your models, logs and key. The empty folder that's left takes no space. If you changed Model location, move that folder too.
4. Use Go to Folder again for ~/Library/Application Support, and find the Ollama folder, which holds your chats. Move it to the Trash only if you don't want them.
5. Do the same for the smaller folders in Ollama's list: ollama and com.electron.ollama in ~/Library/Caches, com.electron.ollama in ~/Library/WebKit, and the app's window state, com.electron.ollama.savedState, in ~/Library/Saved Application State.
6. Open Apple menu › System Settings, click General, then Login Items & Extensions, and check that Ollama no longer appears.
7. Empty the Trash when you're sure.

What about the command-line link?

When Ollama first started, it asked to put a link to its command-line tool in /usr/local/bin. The link is tiny, and once the app is gone it points nowhere. You can leave it, or open that folder with Go to Folder and move the ollama link to the Trash, which may ask for an administrator's password.

If you installed it with Homebrew

Remove it the way you installed it, through Homebrew. A plain Homebrew uninstall leaves ~/.ollama and your models in place, so check for them in Finder as above. The general approach to leftovers is in uninstalling a Mac app and what it leaves behind.

Choosing which models to keep

Removing a library model costs only a download. Removing one you made can cost the work of making it again. Sort your models by where they came from before you remove anything.

A quick way to decide

If the model is	Then
From Ollama's library, and unused for weeks	Remove it; one download brings it back
A second name for a model you keep, with the same ID	Removing it frees nothing; remove it only to tidy the list
A larger or smaller build of a model you also keep	Keep the build you use; the others are full copies
One you created from a Modelfile	Save the Modelfile and its source files first
Imported from a GGUF file you no longer have	Keep it, unless you're sure; Ollama can't download it again
Used by another app, such as an embedding model	Check that app first; it may need the model to work

Mistakes to avoid

• Moving the whole .ollama folder when you meant the models. It also holds your key and logs. Move only models, unless you're uninstalling.
• Deleting the largest blobs in Finder. They're the weights of models you may still use, and other models may share them.
• Changing Model location and expecting your models to follow. They stay in the old folder, taking the same space, until you move them.
• Moving files while Ollama is running. Quit it first, from its menu bar icon, so it isn't writing to the folder while you change it.
• Pasting cleanup commands from a web page. A command that deletes the wrong folder can't be undone, and nothing here needs one.

Keeping it from growing back

Check a model's size in Ollama's library before you pull it, and remove the ones you only tried. Keep large models on an external drive with Model location if your startup disk is small. Recheck the models folder's size with Get Info now and then, since models stay until you remove them.

How the app does it

Select Ollama's models folder in the map, the app's home screen, and it says what the folder is, that Ollama made it, and how its models are removed. The explanation comes from the catalog's entry for Ollama models, which is written and checked by hand, with its sources.

• ~/.ollama/models

  Ollama models

  Language models you've pulled or created with Ollama, stored as shared layers.

  Review first · Explained only: never removed

The folder is Review first: it never starts checked and never counts toward the safe-to-remove total. The app doesn't remove Ollama's models at all. Only Ollama knows which layers each model still needs, so the app explains the folder and leaves removing models to Ollama, as described above.

Ollama's logs and other small folders aren't in the catalog. For items like those, the inspector can give a Best guess, written on your Mac by Apple Intelligence and labelled as a guess, and the app never offers them for removal.

Models arrive in large steps, and the menu bar item measures the folders the catalog knows, AI models included, once a day. What Changed shows which places grew each day, and an alert can tell you when something grows by 10 GB or more in a day, such as a new large model. Both are PRO.

Seeing what Ollama's models take, and why, is free.