Disclosure: Power and Pantry is reader-supported. If you buy through links on this page, we may earn a commission from Amazon or another program at no extra cost to you. See the About page for how this site is funded.
Best Offline AI Apps for Your Phone (2026)
The short answer
Two apps in this comparison genuinely do the thing, and one platform feature does half of it. PocketPal and Google AI Edge Gallery both download an open model onto the phone and run it there, and both developers state in their own words that no internet is required once the model is on the device. PocketPal is the purpose-built chat app: free, open source, on iPhone and Android. The Gallery is Google's own showcase for on-device models: also free, also on both platforms, and labeled experimental by Google itself. Apple's entry is different in kind. There is no Apple chatbot app to download. Apple Intelligence runs a model on the phone itself for part of its work and sends the rest to Apple's servers, and Apple does not mark which route a given request took. If the test is a chatbot that works in airplane mode, every time, the two downloadable apps pass it. Apple's option passes it only for the requests that stay on the device, and you cannot choose which ones those are.
What "offline" has to mean on a phone
The test is the same one this corner applies to computers, set out on the page about running your own AI when the internet is down: download the model while you have a connection, then turn the radios off and ask a question. If an answer comes back, the phone produced it. Every app on this page is built around that single internet moment. The app itself is small; the model is the download, and it happens in advance, on an ordinary day, or it does not happen at all. An app that streams its answers from a server fails the test no matter how it is marketed, which is why the ChatGPT and Gemini apps are not in this comparison: as the setup page covers, they are windows onto services, and with no connection there is nothing behind the window.
The phone version of the idea runs under tighter limits than the laptop version, and the binding limit is memory. A model has to fit in RAM alongside the operating system and everything else the phone is doing, and a phone has less of it than a laptop and no fan to shed heat. The models that make sense here are the small classes from the setup page's ladder, 1B to 4B, with download sizes to match. Google's published model list for the Gallery puts numbers on the point: its listed models run from about 0.6 GB to about 4.4 GB as files, and Google's own estimated peak memory for them runs from about 2 GB to about 6.5 GB. The peak figure is the honest one to plan around. A model whose peak memory approaches the phone's total RAM is a model that will crawl or be closed by the system, whatever its file size says.
How this comparison was put together
An entry earned its place one way: its developer's own material states that inference runs on the device with no internet required. PocketPal's repository puts it in one sentence: a private AI assistant that runs entirely on your phone, no account, no cloud, no internet required. Google's Play Store listing for the Gallery says all model inferences happen directly on your device hardware and no internet is required. Apple's documentation describes an on-device model and, in the same breath, the server route for harder requests, so Apple is included with that split printed rather than smoothed over. Prices are the store listing prices, checked in November 2026. Specifications come from the developers' repositories, model lists, and documentation, and from Apple's support page for eligibility. Nothing here is ranked by commission. This corner's affiliate anchor is Amazon, for the drives and storage its other pages name; the apps themselves are free, and that relationship is disclosed at the top of the page and again here.
The grid: two apps and one platform feature
| Entry | Runs fully offline? | Platforms | Price, checked Nov 2026 | Models | Device or memory requirement, as published |
|---|---|---|---|---|---|
| PocketPal | Yes. Chat runs on the phone after a one-time model download, in the developer's words. Optional online features exist and are off unless turned on. | iPhone and iPad (iOS 15.1 or later, per the App Store listing) and Android | Free. Optional one-time PocketPal Plus purchase and paid Pals inside the app; chat and models are free. | Your choice: GGUF models from the in-app list or from Hugging Face. The developer names the Gemma, Qwen, Phi, and Llama families. | The developer's Play listing gives guidance, not a hard floor: 6 GB of RAM or more for smaller models, 8 GB or more for the larger ones. |
| Google AI Edge Gallery | Yes for inference, in Google's words. Some optional Agent Skills load from the web when used. | Android (version 12 or later) and iPhone (iOS 17 or later), per Google's repository. A macOS version exists and is outside this page's subject. | Free. Open source under the Apache 2.0 license. Google labels it an experimental beta. | Google's curated list, Gemma family at the center, plus custom models you load yourself. Published file sizes about 0.6 to 4.4 GB. | The operating system floors above. Per model, Google publishes an estimated peak memory, about 2 to 6.5 GB across its listed chat models. |
| Apple Intelligence (not an app) | Partly. Features the on-device model handles work with no connection. Requests Apple routes to Private Cloud Compute need one, and Apple does not label the route per request. | Built into the operating system on eligible iPhones; there is nothing to download from a store. | Included with the operating system. Apple states no per-request cost for features built on its framework. | Apple's own on-device model only. No model choice, and no way to load another one. | Apple's support page lists iPhone 15 Pro and Pro Max, iPhone 16 models or later, and iPhone Air. The system reserves up to 8 GB of storage on most eligible models, up to 14 GB on the newest tier. |
Matched offers are coming. This site is new, and the affiliate links for this corner are still being set up. When they are live, this box will hold current offers for the hardware this corner's pages name, such as external drives for storing models and offline libraries, with prices re-checked at that time. The disclosure at the top of this page explains how the site earns.
PocketPal: the purpose-built chat app
PocketPal is the closest thing on a phone to the computer setup in this corner: a chat app whose whole design is running a model you downloaded, on hardware you own, with nothing sent anywhere. It is an open source project, published under the MIT license, and its repository describes the flow in three steps: install the app, download a model, load it and start chatting, at which point you are running AI fully offline. Its App Store listing walks the same path, download the model weights once, then chat offline. Its Play Store listing is blunter: complete offline operation, airplane mode, no signal, underground, anywhere.
The model choice is the widest of the three entries. PocketPal runs GGUF models, the same file format the computer pages in this corner are built around, and it can pull them from its own list or search Hugging Face from inside the app, so the Llama 3.2 sizes from the setup page, 1.3 GB for the 1B and 2.0 GB for the 3B as Ollama lists the weights, are the class of file you are choosing among. The developer's own sizing guidance, from its Play listing, is 6 GB of RAM or more for smaller models and 8 GB or more for the larger ones, with flagship and recent mid-range phones named as the hardware that runs models well. The app also carries a built-in benchmark, so the check this corner recommends, try it on your own machine, is a feature the developer ships rather than homework.
The price structure is plain. The app is free on both stores, and the developer states that chat, models, voices, and tools are free with no account. The App Store lists an optional one-time purchase, PocketPal Plus, at prices from $3.99 to $44.99, and the repository says what it buys: hands-free voice conversation and a local API server, not the chat itself. Premium assistant personas, called Pals, are listed at $0.99 to $2.99. Two honest qualifications belong in this section rather than the fine print. The app includes optional features that do use the network, web search and a remote server mode among them, and the developer's privacy note says they stay off unless you turn them on. The same note lists the app's other network uses, which include fetching device-compatibility rules from a content network at launch. None of that attaches your chats, per the developer, and the App Store's privacy label for the app reads Data Not Collected. But "the app never touches the internet" is a stronger sentence than the developer's own note supports, so this page does not print it. The claim that matters, the chat runs on the phone, is the one the developer makes and the one the airplane mode test settles.
Google AI Edge Gallery: Google's own showcase
The Gallery is Google's demonstration of its own on-device stack, and Google describes it in those terms: a gallery that showcases on-device machine learning and generative AI use cases and lets people try models locally. Its core promise is printed on its Play Store listing as a heading, 100% on-device privacy, with the sentence this page's standard requires: all model inferences happen directly on your device hardware, and no internet is required. The app is free, the project is open source under the Apache 2.0 license, and Google labels the release an experimental beta, with the Play listing adding that the app is in active development. That label is part of the facts, and it cuts both ways: you are getting Google's own engineering for its own models, in software Google itself calls unfinished.
Where PocketPal leaves the model choice open, the Gallery curates it. Google publishes a model list with the file size and an estimated peak memory for each entry, and the chat models on it, checked in November 2026, are these: Gemma3 1B at about 0.6 GB with an estimated peak memory of about 2 GB; Qwen2.5 1.5B at about 1.6 GB and about 2.5 GB peak; Gemma 3n E2B at about 3.1 GB and about 5.5 GB peak; and Gemma 3n E4B at about 4.4 GB and about 6.5 GB peak. Read the peak column against your phone's total RAM before the size column: the largest model on the list wants most of an 8 GB phone's memory at its peak. The app also lets you load custom models of your own, and it carries features beyond chat, image questions, audio transcription, and a prompt laboratory among them, all described by Google as running on the device. One qualification mirrors PocketPal's: the Gallery's Agent Skills can load skills from a URL and use Wikipedia for fact grounding, and a skill that fetches is a feature that needs the network when you use it. The chat inference itself is the on-device part, and that is the part this page is grading.
The practical difference between the two apps is intent. PocketPal is a chat app you settle into, with your choice of model and your own personas. The Gallery is a sandbox Google maintains to show what its models and runtimes can do on your hardware, which makes it the better instrument for answering "what can my phone actually run" and the less settled home for a daily assistant. Both conclusions come from the developers' own descriptions of their software, not from use, and the data basis note at the top of this page stands.
Apple: the option that is not an app
Apple is in this comparison because it is the option most iPhone owners already have, and it needs describing exactly, because it is not the same kind of thing as the two apps above. Apple does not offer a chatbot app. Apple Intelligence is a set of features built into the operating system and its apps: writing tools, summaries, a more capable Siri, and similar help that appears inside the software you already use. Underneath those features sit two routes. One is a model that runs on the device itself. The other is Private Cloud Compute, Apple's servers, which exist precisely for the requests the on-device model cannot carry. Apple's developer documentation states the division in one sentence: on-device models excel at tasks like summarization and refinement, and when you need more reasoning capability and context size, you use Private Cloud Compute or another server model. A request that takes the server route needs a connection by definition. Apple does not show you which route a given request took, and there is no setting that confines Apple Intelligence to the device.
The framework underneath deserves its own sentence, because it is the part of Apple's offering that is genuinely and completely offline. The Foundation Models framework lets any app use the same on-device model that powers Apple Intelligence, and Apple's product page states that features built on it work offline and at no cost per request. That is a developer tool, not a consumer app: it is how other people's apps get an on-device model without shipping one. A phone owner meets it in two places, Apple's own features, and third-party apps built on the framework, whose offline behavior is then that app's design decision. If an app on your iPhone advertises Apple Intelligence inside, the on-device part of what it does will work in airplane mode. The general chat window Apple itself never shipped will not appear, because it does not exist.
Eligibility is Apple's, and it is a hard gate. Apple's support page, checked in November 2026, lists iPhone 15 Pro and iPhone 15 Pro Max, iPhone 16 models or later, and iPhone Air; on the tablet side, iPad mini with the A17 Pro chip and iPad models with M1 or later. An ordinary iPhone 15 is not on the list, and no update adds it. The support page also states the storage the system takes: up to 8 GB on device for most eligible models, and up to 14 GB on the newest tier with 12 GB or more of unified memory. The price is the operating system itself, with two footnotes from the same support page. Features that rely on server-side models carry daily usage limits, and Apple says expanded access to those features will be available for a fee in the future. Those limits and that fee concern the server route. The on-device model is simply there, which is the most that can be said for it, and for the outage use this corner cares about, the honest summary is this: some of what Apple Intelligence does will work with the radios off, some will not, and Apple keeps the sorting to itself.
What a working setup takes
The arithmetic below is the developers' published figures, added up in the open. The pattern is the same for all three entries: storage for the model, memory to hold it, and one download in advance.
- PocketPal: the app itself, then one model file. At the sizes this corner uses, that is 1.3 GB for a 1B model or 2.0 GB for a 3B, as Ollama lists those weights, and the developer's guidance is 6 GB of RAM or more for the smaller models, 8 GB or more for the larger. Total new storage: under 3 GB for the sensible phone setup.
- Google AI Edge Gallery: the app, then one model from Google's list, 0.6 to 4.4 GB as a file. The figure to check first is Google's estimated peak memory for the model you pick, up to about 6.5 GB for the largest listed chat model, against your phone's total RAM.
- Apple Intelligence: no download of yours to manage and no model to pick. The system reserves up to 8 GB of storage on most eligible phones, 14 GB on the newest tier, per Apple's support page, and eligibility is decided by the phone you already own.
Battery is the figure nobody in this comparison publishes, and it belongs in a setup section because it shapes how the tool gets used. Running a model is the hardest sustained work a phone does, and a phone battery is small next to everything else in this site's power pages. The planning rule from this corner's power page applies unchanged: count the hours, not the questions. An evening of asking is a modest load. A phone left awake running an assistant around the clock is a battery decision, and during an outage it is a decision against the same battery that carries your calls, your light, and your radio apps.
What we could not verify
- Our own use of any entry here: we have not installed PocketPal or the Gallery, and we have not run the airplane mode test on a phone ourselves. Every offline claim on this page is the developer's stated claim, quoted at the strength the developer states it, and the test that would settle it on your phone is described above.
- Battery drain: neither PocketPal nor Google publishes a battery figure for chat use, and Apple publishes none for Apple Intelligence. This page prints no per-hour battery number for any entry, because no primary source offers one.
- A hard minimum device for PocketPal: the developer publishes guidance (6 GB of RAM or more for smaller models, 8 GB or more for larger ones) and ships an in-app benchmark, but no single minimum phone or RAM figure appears in the repository or the listings we checked. The grid prints the guidance as guidance.
- Item-for-item parity between the Gallery's Android and iPhone versions: Google's repository links both stores and states both operating system floors, and the published model list we read is written around Google's Android runtime. We did not verify that the iPhone version offers the identical model list.
- Which Apple Intelligence requests stay on the device: Apple describes the two routes and which kind of task suits each, but publishes no per-feature routing table, and its system shows the user no indicator. No source we would print lets us draw that line feature by feature.
- PocketPal Plus's exact price for a given buyer: the US App Store lists the one-time purchase at several prices from $3.99 to $44.99, and the repository describes it as a single optional purchase. We print the range as listed and the repository's description of what it buys.
Questions
Can ChatGPT work without internet on my phone?
No, and the phone changes nothing about that answer. ChatGPT runs on OpenAI's servers, the app is a window onto the service, and there is no model file to download. The same holds for the Gemini app and the other hosted chatbots. Working without internet means a different, smaller model running on the phone itself, which is what both apps in this comparison do.
How do I know an app is really offline and not quietly using the cloud?
Run the test from the top of this page: download the model, switch on airplane mode, and ask a fresh question. An answer in airplane mode was produced on the phone. Beyond the test, prefer developers who state the architecture plainly. PocketPal's code is public, so its claim can be read rather than trusted, and Google states the on-device design of the Gallery in its listing. The harder case is a platform feature like Apple Intelligence, where both routes exist inside one name and the system does not label them; there, the honest position is the one this page takes, partly offline, route not shown.
Which phone do I need?
For the two apps, the answer is memory more than model year. PocketPal's developer asks for 6 GB of RAM or more for smaller models and 8 GB or more for larger ones, and Google's list asks you to match a model's estimated peak memory against your phone's RAM. The Gallery also sets operating system floors, Android 12 and iOS 17. For Apple Intelligence the answer is a list, not a specification: iPhone 15 Pro or Pro Max, iPhone 16 models or later, or iPhone Air, per Apple's support page. A phone outside that list gets nothing from the Apple route, whatever its memory.
How much storage will this take?
One model is the whole cost for the apps: roughly 1 to 2 GB for the small models that suit phones, up to about 4.4 GB for the largest model on Google's list. Apple's route reserves up to 8 GB on most eligible phones and up to 14 GB on the newest tier, and that space is the system's to manage, not a file you pick. The pairing from this corner's offline library page fits beside any of these: the library is files, the phone AI is a program, and the same drive arithmetic applies.
Can the phone AI read my offline library for me?
Not directly, and the seam is the same one the library page describes for computers. The library's package files are read by their reader app, and the chat apps take text you bring to them, so you find the article and paste the passage, within the small context window a phone model has. Software that manages your own documents as a collection the AI can search is a separate job, and the next page in this corner takes it up.
Will it drain my battery?
Faster than idle, by an amount no developer in this comparison publishes, so this page prints no figure. Inference is heavy, sustained work for a phone processor, and the sensible way to budget it is the power page's rule: the load is the hours. Short sessions for specific questions are cheap. Leaving a model loaded and the screen awake is the expensive pattern, on a phone even more than on the laptops the power page covers.
A note on responsibility
A phone model is the smallest class of model in this corner, and it carries the same warning as the largest: it produces fluent, confident text that is sometimes wrong, with no signal attached to the wrong parts. Do not take medical, electrical, structural, or safety instructions from it. During an actual emergency it knows nothing about your situation, because its knowledge froze before the event and it has no way to receive word. For what is happening, the tools are official alerts and the radios in this site's emergency radio comparison and off-grid communication explainer. Use the phone AI the way this corner uses every local model: to explain, summarize, draft, and translate information you already hold, checked against a reference that carries responsibility for its facts. This page's data basis is stated at the top and stands: the developers' published specifications, listings, and documentation, checked in November 2026, with the gaps named where a reader would otherwise assume a test or a number exists.