How to Run a Local LLM on iPhone, Fully Offline

To run a local LLM on an iPhone, install an app that runs open models on the device, download a model that fits your iPhone’s memory, and chat. After the download, it works in airplane mode. Any iPhone from the last several years can run a small model. An iPhone 15 Pro or newer, with 8 GB of RAM, can run 9-billion-parameter models that give noticeably better answers.

You don’t need a Mac, a developer account or any command line. The whole thing takes about ten minutes, most of it spent downloading.

Which iPhones can run a local LLM? #

The limit is RAM. Apple doesn’t advertise it, but the pattern is well established:

Older iPhones with 3 to 4 GB can still run the smallest models, but answers are simpler. Storage matters too: models take roughly 1 to 6 GB each.

Step by step: your first local model on iPhone #

1. Install a local AI app from the App Store. Options are listed below.
2. Connect to Wi-Fi. Model files are large.
3. Pick a model sized for your phone. Start with a 3B or 4B model unless you have an 8 GB iPhone.
4. Download it. A 4B model is roughly 2.5 to 3 GB.
5. Turn on airplane mode and ask a question. If it answers, it’s running entirely on your iPhone.
6. Try a harder question and watch the speed. If replies crawl, switch to a smaller model.

Which model should you start with? #

• 4 GB iPhones: Gemma 4 E2B or Ministral 3 3B (about 2 GB each). Qwen 3.5 0.8B is the fallback for very tight memory; it’s fast but basic.
• 6 GB iPhones: Qwen 3.5 4B. It’s the best all-rounder at this size and supports a thinking mode for math and logic.
• 8 GB iPhones: Qwen 3.5 9B for the best general answers, or GLM 4.6V Flash for reading documents, screenshots and photos.

For how these compare, see the best open-source LLMs for phones, and for the memory maths behind the table, how much RAM you need to run an LLM on a phone.

Apps that run local LLMs on iPhone #

Personal LLM needs iOS 15.1 or later. It reads your iPhone’s actual RAM and shows “Fits your device” (at least 2 GB of headroom), “Should run” (meets the minimum) or the RAM a model needs, so you don’t have to guess from a table. It runs models through llama.cpp using Metal on the iPhone’s GPU. Every reply shows a live tokens-per-second readout, and there’s a built-in benchmark to compare models on your phone. You can switch models from the chat header mid-conversation. The app is free; ads appear only when you’re online, and a one-time purchase removes them.

Isn’t Apple Intelligence already a local LLM? #

Partly. Apple Intelligence runs a roughly 3-billion-parameter model on the device for features like Writing Tools and notification summaries, and it sends larger requests to Apple’s Private Cloud Compute servers (Apple Machine Learning Research). It’s available on the iPhone 15 Pro and iPhone 16 models or later (Apple Support).

The differences from running your own model:

• You can’t pick the model. Apple chooses it and updates it with iOS.
• It isn’t a general chatbot. It powers specific features rather than an open-ended conversation. Siri can hand questions to ChatGPT if you enable that extension, but that goes to OpenAI.
• Some requests leave the phone, to Private Cloud Compute, even if Apple designed it with strong privacy protections.
• Older iPhones don’t get it at all, while a local LLM app runs on phones back to the iPhone 13 and earlier.

What about Core ML, MLX and llama.cpp? #

These are the engines apps use under the hood, and you don’t need to choose one yourself. Core ML is Apple’s framework for running machine-learning models, used widely for vision and small models. MLX is Apple’s array framework for Apple silicon. llama.cpp is an open-source engine built for LLMs, with Metal support on iPhone, and it’s what most apps that load GGUF files use. If an app lets you add models from Hugging Face in GGUF format, it’s almost certainly llama.cpp.

Tips for better speed and battery on iPhone #

1. Close heavy apps before a long session so the model has memory to work with.
2. Leave thinking mode off for casual chat. It’s slower, and worth turning on only for math, logic and code.
3. Low Power Mode slows generation. Turn it off when you want speed.
4. Expect warmth. Long answers from a 9B model will heat the phone. If it throttles, a smaller model is often faster overall.
5. Keep conversations focused. Very long chats fill the context window and slow each reply.
6. Delete models you don’t use to reclaim storage.

What can you do with it offline? #

Plenty: draft and rewrite messages, summarize a PDF (see summarizing a PDF offline), explain a concept, brainstorm, translate phrases, and with a vision model, point your camera at a menu or sign and ask what it says. What it can’t do is browse the web or know today’s news.

Frequently asked questions #

Can an iPhone run an LLM without internet? #

Yes. Once you’ve downloaded a model inside a local AI app, it runs entirely on the iPhone’s chip and works in airplane mode. The internet is only needed for the initial download.

How much storage does a local LLM need on iPhone? #

Between about 0.8 GB for the smallest models and about 6 GB for 9B models. Adding image understanding to a model takes another 200 MB to 1 GB. Keep some free space beyond that for iOS itself.

Can I run a local LLM on an older iPhone? #

Usually, as long as it runs a supported iOS version. Personal LLM needs iOS 15.1 or later. iPhones with 3 to 4 GB of RAM are limited to the smaller models, which answer more simply but still work offline.

Is running an LLM bad for my iPhone’s battery? #

It uses a lot of power while generating, similar to a demanding game, and the phone gets warm. Short chats barely register. Long sessions with big models drain the battery noticeably, so plug in if you’re working through a lot of text.