Partly. Apple Intelligence runs a language model on the iPhone itself, and features built on it, such as notification summaries and many Writing Tools actions, are designed to run on the device. Requests that need a bigger model go to Apple’s Private Cloud Compute servers, and anything you hand to ChatGPT goes to OpenAI; both need a connection. Apple Intelligence also only runs on the iPhone 15 Pro and newer models, so older iPhones need a different route to offline AI.
Why is AI moving from the cloud onto phones? #
For years, assistants sent everything to data centers because phones couldn’t run useful language models. Two things changed. Phone chips gained enough memory and GPU and neural-engine power to run models with a few billion parameters, and model makers learned to shrink models through quantization and distillation with modest quality loss.
The result is a hybrid pattern that Apple, Google and Samsung all now use: a small model on the phone handles quick, private tasks, and a larger model in the cloud handles the rest. Apple’s version of the cloud half is Private Cloud Compute.
Which Apple Intelligence features work offline? #
Apple doesn’t publish a feature-by-feature offline list, and the system decides per request whether the on-device model is enough. This is the general picture:
| Feature | Where it runs | Works in airplane mode? |
|---|---|---|
| Notification summaries | On the device | Yes |
| Writing Tools (proofread, rewrite, summarize) | On the device for most requests; larger ones can use Private Cloud Compute | Usually, for short text |
| Genmoji and Image Playground | Described by Apple as on-device | Generally yes, once set up |
| Clean Up in Photos | On the device | Yes |
| Siri answers that need information (weather, web, sports) | Apple and partner servers | No |
| Siri or Writing Tools requests passed to ChatGPT | OpenAI | No |
| Visual Intelligence lookups via Google or ChatGPT | Those services | No |
| Third-party apps using Apple’s Foundation Models framework | The on-device model | Yes, if the app is built that way |
The quick test is the practical one: switch on airplane mode and try the feature you care about.
How can you see what was sent to Apple’s servers? #
On recent versions of iOS, go to Settings > Privacy & Security > Apple Intelligence Report. It lets you export a log of requests that were sent to Private Cloud Compute over a recent period. If you mostly see short summaries and proofreading staying local, that matches Apple’s design.
What is Private Cloud Compute? #
Private Cloud Compute is Apple’s server system for Apple Intelligence requests that are too large for the phone. Apple says the servers run on Apple silicon, use your data only to fulfil the request, don’t store it afterwards, and publish their software images so outside researchers can check those claims. Apple’s security write-up explains the design.
It’s a far more private design than a typical cloud chatbot. It still means your request leaves the phone, and it doesn’t work without a connection.
Does Apple Intelligence send data to OpenAI? #
Only if you turn on the ChatGPT extension and a request is routed to it. Siri and Writing Tools ask before sending something to ChatGPT, unless you’ve turned that confirmation off. You can use it without a ChatGPT account, and you can leave the extension disabled entirely in Apple Intelligence & Siri settings. If you do sign in to a ChatGPT account, OpenAI’s own data policies apply to what you send.
Which iPhones support Apple Intelligence? #
Apple Intelligence requires an iPhone 15 Pro or 15 Pro Max, any iPhone 16 model including the 16e, or an iPhone 17 model or iPhone Air. All of them have at least 8 GB of RAM. The iPhone 15 and 15 Plus (6 GB), iPhone 14 and earlier don’t get it at all.
Can an older iPhone run AI offline? #
Yes, just not Apple’s. Open models from other labs run through apps that bring their own model. Personal LLM supports iOS 15.1 and later, uses the iPhone’s GPU through Metal, and shows a “Fits your device” badge read from your actual RAM before you download anything. After a one-time download, it works in airplane mode and your chats never leave the phone.
Here’s what each iPhone RAM tier can run from its catalog. “Fits” means at least 2 GB of headroom over the model’s minimum; “should run” means it meets the minimum.
| iPhone (RAM) | Fits | Should run |
|---|---|---|
| iPhone 13, 13 mini (4 GB) | Qwen 3.5 0.8B | Qwen 3.5 4B, Gemma 4 E2B, Ministral 3 3B |
| iPhone 13 Pro, 14, 14 Pro, 15, 15 Plus (6 GB) | Qwen 3.5 4B, Gemma 4 E2B, Ministral 3 3B | Gemma 4 E4B, Qwen 3.5 9B |
| iPhone 15 Pro, 16 family, 17 (8 GB) | Qwen 3.5 9B, Gemma 4 E4B and everything smaller | GLM 4.6V Flash |
| iPhone 17 Pro, 17 Pro Max, iPhone Air (12 GB) | Every model in the catalog |
Qwen 3.5 4B is the recommended first download on most of these. See which phones run local AI best for more detail on RAM and speed.
Apple Intelligence or an open-model app: which should you use? #
They do different jobs, and many people use both.
| Apple Intelligence | Open-model app (Personal LLM) | |
|---|---|---|
| Where it lives | Built into iOS, works across apps | Inside one app |
| Models | Apple’s own | Qwen 3.5, Gemma 4, Ministral 3, GLM 4.6V Flash, or any GGUF by URL |
| Offline | Some features | Everything, after the model download |
| iPhones | 15 Pro and newer | iOS 15.1 and later |
| Photo questions | Visual Intelligence, with lookups via Google or ChatGPT | On-device vision for every catalog model |
| Documents | Summaries via Writing Tools | Attach a PDF, text or Markdown file and ask questions |
| Choice of model and settings | None | Presets, temperature, context size, system prompt per chat |
| Cost | Free with the iPhone | Free, ad-supported; one-time Remove Ads purchase |
Apple’s features are best for quick, system-wide touches like tidying a message or summarizing a notification stack. A chat app with open models is better when you want a conversation, a choice of model, answers about a photo or document with no connection, or AI on an iPhone that Apple doesn’t support. For the notification side of Apple Intelligence, including the accuracy problems, see AI notification summaries. For how Apple compares with other assistants on privacy, see the most private AI assistant.
Frequently asked questions #
Does Apple Intelligence work without internet? #
Some of it does. Features that run on the on-device model, such as notification summaries and many Writing Tools actions, work offline. Requests that go to Private Cloud Compute or ChatGPT, and anything needing live information, don’t.
Which iPhones get Apple Intelligence? #
The iPhone 15 Pro and 15 Pro Max, every iPhone 16 model including the 16e, and the iPhone 17 models and iPhone Air. The standard iPhone 15 and older iPhones aren’t supported.
Is Apple Intelligence private? #
It’s one of the more private designs. On-device requests never leave the phone, and Apple says Private Cloud Compute doesn’t store your data and can be checked by outside researchers. Requests you send to ChatGPT are handled by OpenAI instead.
Can I run an AI chatbot offline on an older iPhone? #
Yes. Apps that run open models on the phone work on iPhones Apple Intelligence doesn’t support. Personal LLM needs iOS 15.1 or later, and small models run with 3 GB of RAM, so even an iPhone 13 can run Qwen 3.5 4B.