The cleanest way to find out whether an iPhone AI app is actually "offline" is to turn on Airplane Mode and ask it something. The ones that answer are on-device. An app that fails with a network error in airplane mode needs the network for a chat. An app that replies "I need an internet connection to help with that" is a wrapper around a cloud service.
Below are ten popular iPhone AI apps and what the airplane-mode test will show you for each, based on where each one actually runs its model. That is a published property of every app here, not a matter of opinion: an app that sends your prompt to a server cannot answer without a network, and an app that runs the model on your phone cannot need one. PocketLLM is first because it was designed from day one to pass this test.
The whole point of this test is that you do not have to take anyone's word for it, including ours. It takes about a minute per app and the instructions are at the bottom.
PocketLLM is launching soon. Private, on-device AI, starting on iPhone, with more platforms planned. No account required, no in-app tracking, and chat processing stays on your device. Join the launch list and be first in.
The test
Install the app, let it finish any first-run model download, then turn on Airplane Mode with Wi-Fi and Cellular off and ask it something substantial. A good prompt is "Write a three-sentence summary of the French Revolution," followed by something harder like "Explain quantum entanglement to a 10-year-old." Note whether it answers at all, and whether the answer is coherent rather than a canned apology.
Two things make this test decisive. It cannot be faked by marketing copy, and it produces the same result on your phone as on anyone else's, because it is testing architecture rather than intent. What follows is what you should expect from each app and why, so that a surprising result tells you something has changed.
The 10 apps, ranked
1. PocketLLM (on-device by design)
Designed from day one for airplane mode. Runs Llama 3.2 1B and other models directly on-device. Because the model runs on the phone, Airplane Mode changes nothing about how it behaves. There is no network path to lose. Currently in the pre-launch waitlist, so you can run this test yourself the day it ships.
2. Private LLM (on-device)
Native iOS app, fully on-device, paid. It works the same way in Airplane Mode as it does online, which is the point: with the weights already on the phone, inference needs no network call.
3. LLM Farm (on-device)
Open-source, free, fully on-device. Requires more setup, since you pick the model and manage quantization yourself, but once a model is downloaded the app has no reason to reach the network at all. Being open source, this is the one app on the list where you can confirm that by reading the code rather than trusting the test.
4. MLC Chat (on-device)
Research-grade, from the MLC-LLM project. Free, open source. The UI is less polished than the top three, but it compiles the model to run on the device itself, so its Airplane Mode behaviour is the same as the others by construction.
5. Apple Intelligence (on iPhone 15 Pro+): on-device, with a cloud fallback
Apple Intelligence runs on-device for simpler requests and tries to fall back to Private Cloud Compute for harder ones. In airplane mode, the on-device path works — summarization, writing tools, simple Q&A. Requests that need Private Cloud Compute need a network connection, so expect them to be unavailable offline. For "some things work, some don't" airplane use, it's fine; for "assume offline for everything," the other four are more predictable.
6. ChatGPT (cloud)
ChatGPT runs its models in the cloud, so a prompt in airplane mode cannot get an answer.
7. Claude (cloud)
Claude also runs in the cloud, like ChatGPT, so a prompt in airplane mode cannot get an answer.
8. Google Gemini (cloud on iOS)
Network error. No on-device fallback (despite Google shipping Gemini Nano on some Android devices; on iOS the Gemini app is fully cloud). It runs in the cloud, like ChatGPT and Claude, so it needs a connection.
9. The "private AI chat" clones (often cloud, not verified)
Search the App Store for "private AI chat" and a rotating cast of near-identical apps appears, many describing themselves as private or offline. Some of them may be thin clients for a cloud API. We are not naming any of them, both because the names change from week to week and because we have not audited any specific one. The pattern is the point. This is precisely the category where the airplane-mode test earns its keep, because the marketing copy and the architecture are frequently unrelated, and one minute in Airplane Mode settles it before you trust the app with anything sensitive.
10. Character.AI (cloud)
Character.AI on iPhone is a thin client for a cloud service. It runs in the cloud, so it needs a connection. Character.AI does not claim to work offline. Included as a baseline: if a known cloud-only app somehow answered you in Airplane Mode, that would mean your radios were not actually off.
The comparison table
| # | App | Where the model runs | Notes |
|---|---|---|---|
| 1 | PocketLLM | On-device | By design, pre-launch |
| 2 | Private LLM | On-device | Paid |
| 3 | LLM Farm | On-device | Open source |
| 4 | MLC Chat | On-device | Open source, research-grade |
| 5 | Apple Intelligence | On-device + Private Cloud Compute | The cloud path needs a network |
| 6 | ChatGPT | Cloud | Cloud-only |
| 7 | Claude | Cloud | Cloud-only |
| 8 | Google Gemini | Cloud | Cloud-only on iOS |
| 9 | "Private AI Chat" clone | Often cloud, not verified | Not named; check each one yourself |
| 10 | Character.AI | Cloud | Cloud-only by design |
PocketLLM is currently in the pre-launch waitlist.
How to test any app yourself
Before you rely on an iPhone AI app for a flight, test it. Swipe down from the top-right corner of your iPhone, tap Airplane Mode, and ask the app a simple question. If it answers, it's on-device (or has cached responses — which is uncommon but worth knowing). If it errors cleanly, it's cloud-only and will fail on the flight. If it returns something weird — a generic apology, a "let me try again later" message, a fallback response — it's doing something hacky and you should not trust its privacy claims either.
The airplane-mode test is also the single best privacy test. If an app can't answer without the network, then by definition your prompts are transiting someone's server. "Private" and "works in airplane mode" are effectively the same category.
The quick answer
The iPhone AI apps that actually work in airplane mode in 2026 are PocketLLM, Private LLM, LLM Farm, and MLC Chat — the four native on-device apps. Apple Intelligence works partially. The other apps on this list run in the cloud or are not verified. Test any "offline" app yourself with airplane mode before trusting the marketing.
For the full on-device iPhone ranking beyond just airplane mode, see our best on-device LLM apps for iPhone roundup.