- Home
- Run an LLM offline on iPhone
How to run an LLM locally on iPhone, offline
Install a local LLM app, download a model that fits your iPhone's memory once over Wi-Fi, then chat in airplane mode. In Pocket AI Lab, which is free and needs no account, 2B to 4B models answer at 26 to 37 tokens per second on an iPhone Air, and every iPhone on iOS 26.1 can run at least a 1B model.
Updated 2026-10-28. Numbers from Pocket AI Lab's Build test on iPhone Air (A19, 11.5 GB RAM as reported by iOS, iOS 26.6.1), 2026-09-03 to 2026-10-23.
What do I need to run an LLM on my iPhone?
An iPhone on iOS 26.1 or later, a few gigabytes of free storage and Wi-Fi for the first download. Your iPhone's RAM decides how big a model it can run.
- Any iPhone that runs iOS 26.1 works, which means iPhone 11 or SE (2nd generation) and newer.
- About 1B models fit on 3–4 GB iPhones, 2B to 3B on 6 GB, up to 4B and Gemma 4 E2B on 8 GB, and larger vision models on 12 GB. The table below has the pick for each.
- Storage for the model, 800 MB to 4.1 GB for the recommended ones, plus 41.1 MB for the app.
- Apple's own on-device model also needs Apple Intelligence turned on, so an iPhone 15 Pro or newer.
- Internet only for the download. Chat itself needs none.
How do I run an LLM offline on iPhone, step by step?
Four steps, and only the download needs a connection.
Install a local LLM app
Get Pocket AI Lab from the App Store. It's free, 41.1 MB, runs on any iPhone with iOS 26.1 and doesn't ask you to sign up.
Pick a model for your RAM
Open the Models tab. The app has scanned your phone's memory and marks the model it recommends: LFM2 1.2B (GGUF) on 3–4 GB iPhones, LFM2 2.6B (GGUF) on 6 and 8 GB iPhones and Gemma 4 E2B (GGUF, Vision) on 12 GB iPhones.
Download it once
Tap the model to download it from Hugging Face, from 800 MB for LFM2 1.2B (GGUF) to 4.1 GB for Gemma 4 E2B (GGUF, Vision). The app checks free space first and keeps downloading in the background.
Chat in airplane mode
Start a chat and type. Turn on airplane mode and the model keeps answering, because it runs on the iPhone's own chip.
Any public model on Hugging Face works too. Tap ⊕ in the Models tab, paste a repo such as bartowski/Llama-3.2-1B-Instruct-GGUF, and the app shows whether it fits your phone before the download starts. GGUF files from the Files app open the same way.
Which local LLM should I run on my iPhone?
Go by memory. These are the models Pocket AI Lab offers first on each class of iPhone, with what they measured on an iPhone Air.
| RAM | iPhones | GPU memory for the app | Start with | Also good |
|---|---|---|---|---|
| 3–4 GB | iPhone 11, 11 Pro, 12, 12 mini, 13, 13 mini, SE (2nd and 3rd gen) | 2.12 GB | LFM2 1.2B (GGUF) | Qwen 3.5 0.8B (GGUF, Vision) |
| 6 GB | iPhone 12 Pro, 13 Pro, 14, 14 Plus, 14 Pro, 15, 15 Plus | 3.10 GB | LFM2 2.6B (GGUF) | Llama 3.2 3B, Llama 3.2 3B (GGUF), LFM2 1.2B (GGUF) |
| 8 GB | iPhone 15 Pro, 16, 16 Plus, 16 Pro, 16e, 17, 17e | 5.34 GB | LFM2 2.6B (GGUF) | Qwen 3.5 4B (GGUF, Vision), Qwen 3.5 2B (GGUF, Vision), Llama 3.2 3B |
| 12 GB | iPhone 17 Pro, 17 Pro Max, iPhone Air | 8.06 GB | Gemma 4 E2B (GGUF, Vision) | Qwen 3.5 4B (GGUF, Vision), Qwen 3.5 2B (GGUF, Vision), Gemma 3 4B (GGUF, Vision) |
GPU memory is the working set the app measured on each class of iPhone (October 2026). Speed is generation tok/s, quality the score on suite v3. More detail and the full ranking on best local LLM for iPhone. For Gemma 4 sizes, see Gemma 4 on iPhone.
What's the difference between llama.cpp, MLX, Core ML and Apple Foundation Models?
They're different engines for running models on Apple hardware. Start with llama.cpp and GGUF files: in our tests it used the least memory and ran the fastest model. MLX brings more vision models, Core ML was slow at writing, and Apple Foundation Models needs no download at all.
| Engine | Format | Measured on iPhone Air | Good for |
|---|---|---|---|
| llama.cpp | GGUF | Llama 3.2 3B (GGUF) at 26.9 tok/s, 721 MB peak. LFM2 1.2B (GGUF) at 77.3 tok/s, the fastest we measured. | Most chat models, the lowest memory use, and vision through a projector file (Gemma 4, Qwen 3.5) |
| MLX | MLX safetensors | Llama 3.2 3B at 23.9 tok/s (throttled), 2,568 MB peak. | Vision models that also take video frames (Gemma 3, Qwen2-VL) |
| Core ML | .mlmodelc | Qwen2.5 1.5B (Core ML) at 3.0 tok/s generation, ~646 tok/s prompt reading, 39 s load. | Experiments. The catalog model is an older one |
| Apple Foundation Models | built into iOS | Not in our benchmark. | Nothing to download, 4K context. Needs Apple Intelligence, so iPhone 15 Pro or newer |
| whisper.cpp | ggml | – | Turning voice messages into text, 80 MB for tiny.en |
Pocket AI Lab has all five built in and shows which one is generating. Every number is on the benchmarks page.
The same Llama 3.2 3B peaked at 721 MB on llama.cpp and 2,568 MB on MLX. llama.cpp maps the GGUF file from storage, and iOS doesn't count those file-backed pages against the app the way it counts weights MLX copies into memory. On a 6 GB iPhone that difference decides whether a model loads at all.
llama.cpp runs on the iPhone's GPU through Metal. Pocket AI Lab ships it built in, so there's nothing to compile. If you'd rather build it yourself, the llama.cpp repository has a SwiftUI example (examples/llama.swiftui) that runs from Xcode on a Mac.
Does a local LLM keep my chats private?
Yes. The model runs on the phone, so prompts, answers, photos and voice recordings are processed and stored on the device and never sent to a server.
There's no account and no sign-in. Models come from Hugging Face as anonymous downloads, and the app never asks for your token. It does connect to a few hosts, none of which receive your conversations, and the privacy page lists each one and what it gets. The source code is on GitHub under MIT, so you can check all of this yourself.
Why does my local LLM crash, slow down or heat up the iPhone?
Nearly always it's memory or heat. iOS closes any app that goes over its memory limit, and a phone that has been generating for a while gets hot and slows down.
The app closes or the model won't load
The model is too close to the limit iOS sets for one app, and that limit is well below the RAM on the box. On a 6 GB iPhone it's about 3.1 GB of GPU memory. Pick a smaller model, or the GGUF build of the same one. Pocket AI Lab checks each model against your phone before downloading and won't recommend one that only just fits, but an imported model can still sit right at the edge. Starting a new chat also drops the long context of the old one.
Answers are slow
2B to 4B models wrote 26.3 to 36.6 tokens per second on an iPhone Air, which is quick. Bigger ones slow down a lot once the phone heats up, and Llama 3.1 8B lost 66% of its speed within one run and averaged 7.6 tok/s. Use a 1B to 3B model for everyday chat, let a hot phone cool down, and run Model Benchmark in the app to see your own phone's numbers.
The phone gets warm and the battery drains
That's the GPU working at full speed for every token. The app shows live CPU, RAM and thermal readings while it generates, so you can watch it happen. A smaller model, shorter answers and not charging while you chat all keep the heat down.
A download fails
Check free storage and Wi-Fi first. The app checks free disk space before it starts, so a full phone fails right away instead of halfway. Gated Hugging Face repos, the ones where you accept a license on the website, can't be downloaded because the app never asks for your token. A mirror of the same model usually works.
Are there other ways to run AI offline on iPhone?
Yes. Cloud chatbots like the ChatGPT app need a connection for every message, but several apps run models on the phone, and developers can build llama.cpp themselves.
PocketPal AI also runs GGUF models through llama.cpp and works on older iOS versions and on Android. Private LLM is a paid app with Siri and Shortcuts, Locally AI runs MLX models, and Google AI Edge Gallery runs Gemma 4 on Google's LiteRT-LM. Pocket AI Lab is the free, open-source one that runs llama.cpp, MLX, Core ML and Apple's model side by side, checks every model against your phone's real memory limit and publishes measured speed and quality for the models it recommends. It needs iOS 26.1. Price, minimum iOS, engine and source code for each app are on the comparison page.
Questions
Can I run an LLM on my iPhone without internet?
Yes. Download a model once in an app like Pocket AI Lab, and after that chat works in airplane mode, with the model running on the iPhone's own chip.
Is there an offline ChatGPT for iPhone?
OpenAI doesn't make one, and the ChatGPT app needs a connection for every message. Local LLM apps fill the gap with open models such as Gemma 4, Qwen 3.5, Llama and LFM2, which run on the phone with no internet.
What is the best local LLM for iPhone?
The models Pocket AI Lab recommends first are LFM2 1.2B (GGUF) on 3–4 GB iPhones, LFM2 2.6B (GGUF) on 6 and 8 GB iPhones and Gemma 4 E2B (GGUF, Vision) on 12 GB iPhones. They're fast, light and answer well, and the full ranking with speed and quality is on our best local LLM page.
How much RAM do I need to run an LLM on iPhone?
Any iPhone on iOS 26.1 can run a 1B model such as LFM2 1.2B (GGUF), which needs 1.4 GB. 6 GB is enough for 2B to 3B models, 8 GB for Gemma 4 E2B with little headroom, and 12 GB for the larger vision models.
Does llama.cpp work on iPhone?
Yes. It runs on the iPhone's GPU through Metal. Pocket AI Lab has it built in for GGUF models, so you don't need Xcode, and in our tests it ran LFM2 1.2B (GGUF) at 77.3 tok/s.
How much storage does a local LLM need on iPhone?
From 800 MB for LFM2 1.2B (GGUF) to 4.1 GB for Gemma 4 E2B (GGUF, Vision) among the recommended models, plus 41.1 MB for the app. A typical 3B model at 4-bit is about 2 GB.
Does running an LLM drain the iPhone battery?
It uses more power than most apps, because the GPU works hard for every word. Long sessions warm the phone up, and once it gets hot iOS slows it down.
Is Pocket AI Lab free?
Yes. It's free on the App Store with no subscription or in-app purchases, needs no account, and the source code is MIT on GitHub. Models are free downloads, each under its own license.