Some Mac apps download AI models for themselves, without asking you to choose one. The best-known example is the Chrome on-device model on a Mac: when your Mac meets Google's requirements, Chrome downloads Gemini Nano, a language model of about 4 GB, into its own folder in your Library and keeps it up to date in the background.
Browsers, dictation apps and photo and video tools do the same, each in its own place. Deleting their folders in Finder rarely lasts, because the app downloads the model again the next time a feature needs it. The lasting fix is in each app's own settings: turning off the feature, or deleting the model in the app's model list.
Apple Intelligence is the exception. Its models belong to macOS, which decides when they arrive and keeps them in protected folders, and there's no supported way to remove them.
This post goes app by app: what each one downloads, how big it is, where it lives, and which setting removes it for good.
Which Mac apps download AI models by themselves
A model is a large file of learned numbers that a feature needs in order to run on your Mac rather than on a company's servers. Apps rarely include it in the download you install. They fetch it later, once your Mac qualifies and a feature first asks for it.
Why don't apps include the model when you install them?
The developers' documentation gives three kinds of reason:
- Not every Mac can run it. Chrome and Edge check memory, graphics and free space first, and download nothing to a Mac that falls short.
- The model changes more often than the app. Chrome checks for a newer model each time it starts, and replaces the old one.
- Few people use every model. Dictation apps offer a dozen models, and you download only the ones you pick.
So an app can look modest in your Applications folder while its model sits elsewhere, often several times larger than the app itself.
How is this different from a local model runner?
With Ollama, LM Studio or a Hugging Face library, you choose each model, and you know it's there. The guide to local AI models covers where those tools keep them.
The apps in this post decide for you. A browser downloads its model when a website or one of its features asks for it, and a dictation app often downloads one during setup. You may never have seen a progress bar.
Where the models come from, at a glance
| App | What it downloads | Typical size | How to stop it |
|---|---|---|---|
| Google Chrome | Gemini Nano | About 4 GB | Turn off On-device AI in Chrome's settings |
| Microsoft Edge Canary and Dev | Phi-4-mini or Aion-1.0-Instruct | Not published | Off unless you turn on Edge's flags |
| Firefox | Models for translations, tab groups, PDF alt text and more | Listed on its On-Device AI page | Block AI enhancements in AI Controls |
| MacWhisper, Superwhisper, VoiceInk | Whisper, Parakeet and other speech models | 0.1 to 3 GB each | Delete them in the app's model list |
| Draw Things, DiffusionBee | Image and video models | 2 GB to over 50 GB each | Delete them in the app's model manager |
| Apple Intelligence | Apple's language and image models | Up to 8 or 14 GB | macOS manages them; no removal |
The rest of this post takes these in turn, starting with the one most people meet without knowing it.
The Chrome on-device model on a Mac: Gemini Nano
Gemini Nano is Google's smallest Gemini model, made to run on a phone or a computer rather than in a data centre. Chrome uses it for some of its own features, and lets websites and extensions use it through built-in AI interfaces.
What does Chrome use Gemini Nano for?
Google's help page lists help with writing or rephrasing text, scam warnings, summaries of web pages and organizing your tabs. Text you send to these features is processed on your Mac, not on Google's servers.
Not every AI feature in Chrome uses it. Google notes that some of them don't rely on the on-device model, so those keep working whatever you choose here.
What triggers the download?
The first request for it. Google's developer documentation says the download starts the first time something calls one of the interfaces that needs the model, not when Chrome is installed. That can be one of Chrome's own features, an extension you added or a website you visited.
A download that's cut off resumes. Google says it carries on if you close the tab that started it, and picks up again the next time Chrome starts, as long as that's within 30 days.
Which Macs does Chrome download it to?
Google's developer documentation lists the requirements:
- macOS 13 Ventura or later.
- At least 22 GB of free space on the volume that holds your Chrome profile. The help page for users rounds this to about 20 GB.
- Either a graphics processor with more than 4 GB of video memory, or 16 GB of memory and at least four processor cores.
- An unmetered connection, for the first download only.
Chrome also picks a size. Its documentation says it estimates the graphics performance and downloads a larger variant, such as one with 4 billion parameters, or a smaller one with 2 billion, or runs the model on the processor instead.
Where does Chrome keep the model?
In your Library, inside Chrome's own folder: ~/. Inside is a folder named after the model's version, and in it a file called weights.bin, which holds almost all the space.
The folder sits next to your Chrome profiles, not inside them. Chromium's documentation gives each channel of Chrome its own folder, so Chrome Beta, Dev and Canary each have their own OptGuideOnDeviceModel folder, and each can hold its own copy of the model.
| Chrome channel | Its folder in Application Support |
|---|---|
| Chrome | Google/Chrome |
| Chrome Beta | Google/Chrome Beta |
| Chrome Dev | Google/Chrome Dev |
| Chrome Canary | Google/Chrome Canary |
How big is Gemini Nano in Chrome?
About 4 GB in reports from 2026. MacRumors and The Register described a file of around 4 GB, and one author measured 4.27 GB on a Mac. Earlier reports from 2025 put it nearer 3 GB.
Google doesn't publish a fixed size, and says it changes as Chrome updates the model. To see the current size on your Mac, type chrome://on-device-internals in Chrome's address bar. Google's developer documentation points to that page for the model's status and size.
- Whisper tiny0.1 GB
- Parakeet Multilanguage0.5 GB
- Whisper small0.5 GB
- Cohere Transcribe1.3 GB
- Whisper large-v3 turbo1.6 GB
- Whisper large-v33.1 GB
- Gemini Nano in Chrome4.3 GB · Usually safe
- SDXL Base in Draw Things7 GB · Review first
- Apple Intelligence, Apple's ceiling8–14 GB · Managed by macOS
Whisper sizes from whisper.cpp's model list, converted from MiB and GiB; Parakeet and Cohere Transcribe from Superwhisper's model guide; Gemini Nano as measured on a Mac in May 2026 reports, and Google says it changes with updates; SDXL from Draw Things' published download sizes; Apple's figures from its Apple Intelligence support article.
When does Chrome remove the model by itself?
Chrome deletes Gemini Nano without asking in a few situations, which Google's developer documentation lists:
- Free space runs low. If the free space on the volume drops below 10 GB after the download, Chrome removes the model.
- Your Mac stops qualifying for 30 days. For example, if nothing has used the model in that time, or the hardware no longer meets the requirements.
- An enterprise policy turns the feature off, on a Mac your organization manages.
Chrome doesn't download it again on its own after these removals. A feature or website has to ask for the model again, and your Mac has to meet the requirements once more.
Do updates replace the model or add to it?
They replace it. Chrome checks for a new model when it starts, and Google says updates are full downloads, swapped in without interrupting the features. So an update downloads about 4 GB again, rather than a small patch.
Some interfaces also download small add-ons for the base model, called LoRA weights. Chrome checks for updates to those every day, and purges them 30 days after the base model is removed.
How to remove Chrome's model and stop it coming back
Google's help page is clear that the setting is the only supported way to manage these models. It also says deleting the model's files by hand isn't recommended and may not stop future downloads.
Turn off On-device AI in Chrome
- Open Chrome, click the More menu (the three dots at the top right), and choose Settings.
- Click AI Innovations in the sidebar.
- Turn off On-device AI.
Chrome deletes the model's files and doesn't download them again while the setting stays off. Google says the setting applies per user and per device, so set it on each Mac, and in each channel of Chrome you use.
If you don't see AI Innovations, type "on-device" in the search field at the top of Chrome's settings. In May 2026, MacRumors found the switch under System, so its place has moved between versions. Google also notes the setting isn't available on every device.
Why doesn't deleting the folder in Finder last?
Because nothing has told Chrome to stop. The next time a feature or a website asks for Gemini Nano and your Mac qualifies, Chrome downloads the full model again. You'd free about 4 GB for a few days, then lose it again without noticing.
If you've already moved the folder to the Trash, turn off the setting anyway, so Chrome doesn't fetch the model again.
What stops working when you turn it off?
The features that rely on the model: writing help, page summaries, tab organization and the scam warnings that use it. Websites and extensions that use Chrome's built-in AI find the model unavailable, and do whatever their developers planned for that case.
You can change your mind at any time. Turn the setting back on, and Chrome downloads the model again the next time a feature asks for it, as long as your Mac has the free space.
What if your Mac is managed by an organization?
Google's help page says Chrome Enterprise deployments follow an enterprise policy for this setting. If the switch is greyed out, or the model keeps returning on a work Mac, the policy decides, and your administrator controls it.
Does Chrome keep other model files?
Chrome keeps a handful of other folders in its Application Support folder, some of them for small models behind other features. They're tiny next to Gemini Nano, and Google doesn't document them for users. Leave them to Chrome; deleting Chrome's internal folders can lose settings or break features.
Edge and Firefox: other browsers with built-in models
Chrome isn't the only browser that downloads models. Microsoft and Mozilla take different approaches, and each gives you a different way to stop it.
Does Microsoft Edge download an AI model on a Mac?
Only in its test versions, for now. Microsoft documents a built-in model, Phi-4-mini, for Edge Canary and Edge Dev from version 138, as a developer preview. From version 150, those channels can use a smaller prerelease model, Aion-1.0-Instruct, which also runs without a strong graphics processor.
Both need a flag turned on in Edge's experimental settings before a website can use them. If you never turned on those flags, Edge hasn't downloaded either model.
How much space does Edge's model need?
Microsoft's requirements mirror Chrome's: macOS 13.3 or later, at least 20 GB free on the volume with your Edge profile, and 5.5 GB of video memory for Phi-4-mini. If free space drops below 10 GB, Edge deletes the model so other browser features keep room to work.
Microsoft doesn't publish the model's size. To stop a test version from using it, turn the flags back off, which is where you turned them on.
How do you see and remove Firefox's on-device models?
Firefox lists them on a page of their own:
- Type about:addons in Firefox's address bar and press Return.
- Select On-Device AI in the list on the left. If it isn't there, no feature or extension is using an optional model.
- Find the model, click the trash icon, and confirm.
The page shows each model's size and what it's for. Mozilla notes that removing a model leaves the feature in place. Use the feature again, and Firefox downloads the model again.
How do you stop Firefox downloading them again?
Use AI Controls, a part of Firefox's settings since version 148:
- In the menu bar, click Firefox, then choose Settings.
- Select AI Controls on the left.
- Turn on Block AI enhancements, or set single features to Blocked.
Mozilla says that blocking a feature that runs on your device removes any model it already downloaded. The panel covers translations, alt text for PDF images, AI-enhanced tab groups, speech recognition, key points in link previews and the AI chatbot in the sidebar, which runs online.
| Browser | Model | Where to stop it | Removed when space is low? |
|---|---|---|---|
| Chrome | Gemini Nano | Settings › AI Innovations › On-device AI | Yes, below 10 GB free |
| Edge Canary and Dev | Phi-4-mini, Aion-1.0-Instruct | The flags you turned on | Yes, below 10 GB free |
| Firefox | Several models | Settings › AI Controls | Not documented |
Dictation and transcription apps
Apps that turn speech into text on your Mac need a speech model, and many let you choose among a dozen. Each one you download stays until you remove it, so trying a few models leaves several gigabytes behind.
Why do speech models vary so much in size?
Mostly by how many parameters they have. OpenAI's Whisper, which most of these apps offer, comes in sizes from tiny to large. whisper.cpp, the library many Mac apps build on, lists its files from 75 MiB for tiny to 2.9 GiB for each large model.
| Whisper model | Disk size in whisper.cpp |
|---|---|
| tiny | 75 MiB |
| base | 142 MiB |
| small | 466 MiB |
| medium | 1.5 GiB |
| large-v3 | 2.9 GiB |
| large-v3-turbo | 1.5 GiB |
| large-v3-turbo, quantized | 547 MiB |
Quantized versions store each number in fewer bits, so they're much smaller at a small cost in accuracy. Larger models are more accurate but slower. Most people settle on one, and the rest are dead weight.
MacWhisper
MacWhisper keeps its data in your user folder and installs nothing system-wide, its documentation says. Downloaded models go in a Models folder inside ~/, next to its database of transcripts.
The Mac App Store version, Whisper Transcription, is sandboxed, so the same folders sit inside ~/. Models are downloaded and switched in the app's Manage Models window.
Is it safe to delete MacWhisper's models folder?
Not as a whole, because the same Application Support folder holds your transcripts. MacWhisper's documentation says its data can be inspected or removed by hand, but it doesn't describe removing a single model. Start in Manage Models, and if you remove files by hand, quit MacWhisper first and touch only the Models folder.
A model you remove downloads again when you choose it. Your transcripts stay in the database, and don't depend on which model made them.
Superwhisper
Superwhisper offers local voice models such as Cohere Transcribe, 1.3 GB, Parakeet Multilanguage, 494 MB, and Whisper Large v3 Turbo, 1.6 GB. Its guide puts a local setup at a one-time download of 75 MB to 3 GB.
To remove one, open Settings, then the Models library. Superwhisper's guide says you can browse every model there, download local ones and remove the ones you no longer use.
VoiceInk
VoiceInk lists its local models under AI Models › Local, from Whisper and Parakeet to Cohere Transcribe and SenseVoice. Each model has a card, and its documentation says the card's menu deletes the downloaded files or shows them in Finder.
VoiceInk can also import a whisper.cpp model file you already have. An imported file is your own, so keep the original if you might want it back.
What about apps that use Apple's speech models?
They don't keep their own copy. Since macOS 26, apps can use Apple's speech recognition framework, whose models Apple's developer documentation describes as downloaded from Apple's servers and managed by the system.
Once one app downloads a language's speech model, macOS shares it with other apps and keeps it updated. Each app gets a limited number of language reservations, and when an app releases one, macOS removes the model later. That's why MacWhisper's model list shows a dash for their size: macOS manages that storage.
OpenAI's Whisper package
If you've run Whisper from Python, or used a tool built on OpenAI's own package, its models live in ~/, one file per size you've used. The catalog entry, OpenAI Whisper models, notes that the next transcription with a removed model downloads it again first.
Photo, image and video apps
Image and video models are among the largest models on any Mac, and apps that make or enhance pictures download them by the gigabyte.
Image generators
Draw Things keeps its models in its app container, and DiffusionBee in a hidden folder in your home folder. A single model with the files it needs runs from about 2 GB to over 50 GB, and video models are larger still.
Both apps remove models in their own windows. In Draw Things, open Settings, click the Model menu, then Manage. The post on image generation models covers both apps, ComfyUI and the shared files that make deleting by hand risky.
Topaz Video keeps models inside the app
Some apps put the models they download inside their own app package. Topaz Labs documents the default models folder for Topaz Video on a Mac as /. Its preferences show the same place as the Resource Folder.
That changes where the space shows up. Finder's Get Info on the app includes the models, and Storage settings counts them with the app under Applications, so the app can look far larger than the download you installed.
Can you remove models from inside an app?
Don't. Deleting files inside an app's package can stop it working, and Topaz warns that moving the models folder can stop Topaz Video finding its models, which leads to export errors. A reinstall puts back what you removed anyway.
Use the app's own options instead, or remove the whole app if you no longer need it. Deleting the app deletes the models inside it with it.
| App | Where its models live | How to remove them |
|---|---|---|
| Draw Things | ~/Library/Containers/com.liuliu.draw-things/Data/Documents/Models | Settings › Model › Manage |
| DiffusionBee | ~/.diffusionbee/downloaded_assets | Quit the app; move the folder to the Trash |
| Topaz Video | Inside the app, in Contents/Resources/models | Through the app, or remove the app |
| MacWhisper | ~/Library/Application Support/MacWhisper/Models | Manage Models, or the folder with the app quit |
| Chrome | ~/Library/Application Support/Google/Chrome/OptGuideOnDeviceModel | Turn off On-device AI |
Apple Intelligence: the model you can't remove
Apple Intelligence is a useful contrast. It's the largest AI download on many Macs, yet it isn't an app's model at all: it belongs to macOS.
How is Apple Intelligence different?
macOS downloads its models when your Mac is eligible, updates them with macOS, and keeps them in protected folders under /. Apple says Apple Intelligence needs up to 14 GB on Macs with M3 or later and at least 12 GB of memory, and up to 8 GB on other supported Macs.
Storage settings counts that space under macOS, not System Data, and there's no supported way to delete it. The post on Apple Intelligence storage explains why Storage settings can show more than Apple's figure on macOS 27.
Do apps that use Apple Intelligence download their own models?
Not if they use Apple's Foundation Models framework. Apple describes its on-device model as the one that powers Apple Intelligence, so an app built on it shares the copy macOS already keeps. Apple's speech framework works the same way, as above.
So the same kind of feature can cost very different space. A dictation app that uses Apple's speech models adds no model of its own; one that offers Whisper adds up to 3 GB for each model you choose.
What can you change?
On macOS 26 Tahoe, you can turn Apple Intelligence off in System Settings › Apple Intelligence & Siri; macOS manages the models after that, and the space may not come back right away. On macOS 27 Golden Gate, Apple documents switches for single features, but no single off switch.
| Apple Intelligence | Models apps download | |
|---|---|---|
| Who decides | macOS | The app, when a feature asks |
| Where they live | Protected system folders | The app's Library folders, or inside the app |
| Counted in Storage settings as | macOS | Usually System Data or Applications |
| Can you remove them? | No | Yes, in the app |
| Does deleting the folder last? | Not possible | Rarely; the app downloads it again |
| Shared between apps? | Yes | No; each app keeps its own copy |
How to find AI models apps have downloaded
No single screen lists them. Storage settings shows where the space is counted, Finder shows the folders, and each app's own page shows what it's for.
What Storage settings shows
Choose Apple menu › System Settings, click General, then Storage. Models in an app's Library folders usually land in System Data, the category for what Storage settings can't place elsewhere. Models inside an app's package count under Applications.
Storage settings doesn't name any of them. A large System Data figure on a Mac with Chrome, a dictation app and an image app may be partly their models.
Finding the folders in Finder
In Finder, choose Go › Go to Folder, type a path from the tables above, and press Return. Select the folder and choose File › Get Info to see its size. The tilde stands for your home folder, and Go to Folder opens hidden folders too.
To look more widely, open your Library this way and use Finder's search with a File Size condition, set to larger than 1 GB. Model files are among the largest files most Macs hold, and their names, such as weights.bin or ggml-large-v3.bin, often give them away.
Inside each app
The app's own page is the most accurate, because it knows what each file is for:
- Chrome: chrome://on-device-internals shows the model's status and size.
- Firefox: the On-Device AI page in about:addons lists each model and its size.
- Dictation apps: Manage Models in MacWhisper, Models library in Superwhisper, AI Models › Local in VoiceInk.
- Image apps: Manage in Draw Things' model menu.
Why does the number keep changing?
Because these apps act on their own. Chrome can download a model one week, replace it with an update the next month, and remove it when free space drops below 10 GB. A model that vanished by itself can return by itself when space is free again.
That makes a single snapshot misleading. Watching which folders grow, day by day, says more than one look at Storage settings.
Deciding what to keep
These models only cost space; they don't slow your Mac when nothing is using them. So the question is simply whether you use the feature.
A quick way to decide
| If | Then |
|---|---|
| You use Chrome's writing help, summaries or scam warnings | Keep the model; Chrome manages it |
| You never use Chrome's AI features | Turn off On-device AI to remove it and stop it returning |
| You tried several dictation models | Keep the one you use; delete the rest in the app |
| An image app holds models you made or imported | Keep the original files before deleting anything |
| An app keeps models inside its package | Leave them; remove the whole app if you're done with it |
| Free space is tight and you use Apple Intelligence | Look elsewhere; its models can't be removed |
Mistakes to avoid
- Deleting Chrome's model folder and stopping there. Chrome downloads it again the next time a feature asks. Turn off the setting.
- Moving a whole Application Support folder. MacWhisper's holds your transcripts, and Chrome's holds your profiles, with their bookmarks and history. Remove only the models folder, and only when the app's own option is missing.
- Editing inside an app's package. It can break the app, and a reinstall puts the files back.
- Pasting cleanup commands from a web page. A mistyped command can delete the wrong folder, and nothing here needs one.
- Expecting a speed-up. Removing a model frees disk space. It doesn't make your Mac faster.
Keeping them from coming back
Turn off the features you don't use, in each app, rather than deleting folders. For Chrome that's On-device AI; for Firefox, AI Controls; for dictation and image apps, keep only the models you use. After a major app update, check the setting is still off, since apps can add features and move their switches.
Keep at least 10 GB free, too. Below that, Chrome and Edge remove their models by themselves, and macOS updates and your other apps need room too.
How the app does it
Select a model folder in the map, the app's home screen, and the inspector says what it is, which app made it and how that app removes it. The explanations come from the catalog, which is written and checked by hand, with its sources.
~/
Library/ Application Support/ Google/ {Chrome,Chrome Beta,Chrome Dev,Chrome Canary}/ OptGuideOnDeviceModel The Gemini Nano language model Chrome downloads in the background to run AI features on your Mac.
Usually safe · Explained only: the app shows you how
~/
.cache/ whisper Speech recognition models downloaded by OpenAI's open-source Whisper package, one file per model size.
Usually safe · Moves it to the Trash
~/
Library/ Containers/ com.liuliu.draw-things/ Data/ Documents/ Models Image and video generation models downloaded or imported in Draw Things.
Review first · Explained only: the app shows you how
~/
.diffusionbee/ {downloaded_assets,downloads,imported_models} Stable Diffusion models that DiffusionBee downloaded or that you imported into it.
Review first · Moves it to the Trash
/
System/ Library/ AssetsV2/ com_apple_MobileAsset_UAF_FM_* The on-device language and image models that Apple Intelligence features use.
Managed by macOS · Explained only: never removed
Chrome's on-device model is Usually safe: it never starts checked, and the app explains Chrome's own setting, which removes the model and stops it downloading again. Whisper's models are Usually safe too. Draw Things' and DiffusionBee's models are Review first, because some may be imported and impossible to download again. Apple Intelligence is Managed by macOS: explained, never moved.
Folders the catalog has no entry for, such as MacWhisper's or Topaz Video's models, get a Best guess in the inspector, written on your Mac by Apple Intelligence and labelled as a guess, and the app never offers them for removal.
Models arrive by themselves, so What Changed helps here: it shows which places grew each day, and an alert can tell you when something grows by 10 GB or more in a day. Both are PRO, along with moving the folders the catalog allows to the Trash. Seeing what each model is, and where it came from, is free.
Frequently asked questions
What is OptGuideOnDeviceModel on my Mac?
It's the folder where Google Chrome keeps Gemini Nano, the language model behind some of its AI features, in Chrome's Application Support folder. The weights.bin file inside holds almost all of it, about 4 GB in 2026 reports. Chrome downloads, updates and removes it by itself.
Can I delete weights.bin from Chrome?
You can, but it won't last. Chrome downloads the model again the next time a feature or website asks for it. Turning off On-device AI in Chrome's settings deletes the model and stops it coming back.
Why did Chrome download a 4 GB file without asking?
Chrome downloads Gemini Nano when a feature, an extension or a website first asks for its built-in AI, and only on computers that meet Google's requirements, including about 22 GB of free space. Google's help page explains the download and the setting that controls it.
Does turning off On-device AI in Chrome break anything?
Only the features that rely on the model, such as writing help, page summaries, tab organization and some scam warnings. Browsing, sync and Chrome's other features keep working, and you can turn the setting back on at any time.
Do Firefox and Edge download AI models on a Mac?
Firefox downloads models for features such as translations and tab group suggestions, lists them on the On-Device AI page in about:addons, and removes them when you block a feature in AI Controls. Edge downloads its built-in model only in its Canary and Dev versions, after you turn on a flag.
Where does MacWhisper store its models?
In a Models folder inside ~/Library/Application Support/MacWhisper, next to its database of transcripts. The Mac App Store version, Whisper Transcription, keeps the same folders inside its sandbox container. Manage the models in the app's Manage Models window.
Do AI models in apps slow down my Mac?
Not while nothing uses them. A downloaded model takes disk space, and it uses memory only while a feature is running it. Removing one frees space; it doesn't make your Mac faster.
Sources
- How to get the next generation of Apple Intelligence · Apple
- AssetInventory · Apple Developer
- Manage on-device Generative AI models in Chrome · Google
- Get started with built-in AI · Chrome for Developers
- On-device AI models in Firefox · Mozilla
- Choose a model and create your first mode · Superwhisper