Yes, an iPad can run a local LLM, and the recent ones are among the best mobile devices for it. What decides how big a model you can run is memory, and M-series iPads have more of it than most phones: the current iPad Pro with the M5 chip has 12 GB or 16 GB, according to Apple’s specs. Install an app that runs open models, download one on Wi-Fi, and you can chat, ask about photos and question your PDFs with no connection at all.
Can your iPad run an LLM? #
Almost any iPad that runs iPadOS 15.1 or later can run a small model. How good the experience is depends on RAM:
- The model has to fit in memory. A 4B model needs roughly 3 GB, a 9B model 6 GB or more, plus room for the conversation.
- Apple doesn’t show RAM in Settings. It publishes memory for the iPad Pro (the M5 model has 12 GB with 256 GB or 512 GB of storage, and 16 GB with 1 TB or 2 TB). For other models, check Apple’s tech specs, or let the app read it for you.
- The GPU does the work. Apps built on llama.cpp run models on the iPad’s GPU through Metal, so you don’t need to worry about the Neural Engine.
Which model should you run on your iPad? #
| iPad memory | Start with | Also try |
|---|---|---|
| 3 to 4 GB | Qwen 3.5 4B (2.74 GB) | Gemma 4 E2B (2.04 GB) |
| 6 GB | Qwen 3.5 4B | Gemma 4 E4B (3.00 GB) |
| 8 GB | Qwen 3.5 9B (5.68 GB) | GLM 4.6V Flash (6.17 GB) for photos and documents |
| 12 to 16 GB | Qwen 3.5 9B with a larger context | Bigger bring-your-own GGUF models around 12B |
Download sizes are from the Personal LLM catalog. For the thinking behind these sizes, see how much RAM you need to run an LLM and what 4B, 9B and E2B mean.
How to set up a local LLM on iPad #
- Install Personal LLM from the App Store. It runs on iPad as well as iPhone, costs nothing and needs no account.
- Open the Models tab. Every model card shows a “Fits your device” badge read from your iPad’s actual memory.
- Download a model on Wi-Fi. Downloads come straight from Hugging Face, and they pause and resume if the connection drops.
- Start a chat and choose the model from the chat header.
- Add image support if you want to ask about photos (195 MB to 1 GB extra per model).
- Try it in airplane mode. Nothing changes, because nothing needs the network.
Every reply shows a live tokens-per-second readout, and the built-in benchmark lets you compare models on your iPad. If you want a specific model that isn’t in the catalog, paste a link to any GGUF file; see how to run any GGUF model.
What’s an iPad especially good for? #
The bigger screen and keyboard change how you use a local model.
- Documents. Attach a PDF, text or Markdown file from the Files app and ask about it. The text is searched on the iPad and the model answers from the matching passages, citing the ones it used. See how to chat with a PDF offline.
- Longer writing. With a keyboard, drafting and rewriting long pieces is far easier than on a phone.
- Studying. Paste in notes and ask for summaries, quiz questions or simpler explanations.
- Photos and screenshots. Ask about a diagram, a form, a chart or a page of handwriting.
- Travel. On a plane, an iPad with a 9B model is a capable writing and study assistant that needs no Wi-Fi.
Local LLM or Apple Intelligence on iPad? #
Apple Intelligence is available on iPad mini with A17 Pro and on iPads with M1 or later, according to Apple’s requirements page. It’s built into Writing Tools, Siri and other system features, and some requests go to Apple’s Private Cloud Compute servers.
A local LLM app is different in a few ways:
| Apple Intelligence | Local LLM app | |
|---|---|---|
| Works on older iPads | No, M1 or A17 Pro and later | Yes, with smaller models |
| Choose the model | No | Yes: Qwen, Gemma, GLM, Ministral or your own |
| Open-ended chat | Limited | Yes |
| Custom system prompts and settings | No | Yes |
| Fully offline | For some features | Yes, after download |
| Built into other apps | Yes | No, it’s its own app |
Many people use both. More on what Apple’s features do offline is in does Apple Intelligence work offline.
What an iPad can’t do #
- Very large models. Even 16 GB isn’t enough for 70B-class models, which need around 40 GB at 4-bit.
- Current events. A local model has no web access and knows nothing after its training data.
- Cloud-model quality on hard problems. Frontier cloud models are much larger. A 9B model is very good for its size, not a replacement for them on complex research or code.
If you’re considering running Ollama instead, it’s a desktop tool; can you run Ollama on a phone explains the options for mobile devices.
Frequently asked questions #
Can an iPad run an LLM offline? #
Yes. Once you download a model in an app like Personal LLM, it runs entirely on the iPad’s own chip and works with Wi-Fi off. Nothing you type or attach is uploaded.
Which iPad is best for running local AI? #
The one with the most memory you can afford. The current iPad Pro has 12 GB or 16 GB, enough for 9B models with room to spare. Other M-series iPads handle 4B models well and often 9B models too; the app’s fit badge tells you for your exact iPad.
Does a local LLM work on an older iPad? #
Yes, if it runs iPadOS 15.1 or later, but stick to smaller models such as Qwen 3.5 4B or Gemma 4 E2B. Older iPads with less memory will be slower and may close the app if a model is too large.
Is running an LLM on iPad private? #
Yes, when the model runs on the device. In Personal LLM, the network is only used to download models and to show ads, and your chats, photos and documents never leave the iPad.
Can I use the same app on my iPhone and iPad? #
Yes, Personal LLM runs on both. Chats stay on each device, and you can move them with the app’s JSON backup and restore.