Which iPhones Can Run Local AI? A Device Compatibility Guide

The question “will this run on my phone?” turns on one number Apple doesn’t advertise: RAM.

Not storage. Not chip generation. Not model year. A model either fits in available memory or it doesn’t load — and Cloaked checks this before it will let you download anything, so the answer is enforced rather than advisory.

Why RAM is the binding constraint

To run, a model’s weights must be resident in memory in full. There is no partial loading and no swapping to storage, because inference reads essentially every weight for every token generated.

But weights are only part of the requirement, which is why the RAM figures below are much larger than the download sizes:

• The KV cache grows with conversation length and cannot be reclaimed while the conversation is live — see Context Windows Explained.
• The vision encoder and image tensors add roughly 0.5–1GB at peak on multimodal models.
• Heavy turns cost more. Web search plus thinking mode in one turn is the worst case, and the minimum has to cover it.
• iOS reserves a large share of total memory for the system, and other apps hold memory too.

Exceed what the system will grant and iOS terminates the app rather than slowing it down. That is why the stated minimums are conservative — they cover the peak, not the average.

What each model actually requires

These are the enforced figures from the app’s model catalogue, not estimates.

★ Recommended default.

The ratio of RAM to file size is the thing to notice. Qwen 3.5 4B is a 2.9GB download that requires 10GB — because it is multimodal, has a 32K context, and the minimum has to survive a long conversation with an image attached.

How much RAM your device has

Apple doesn’t publish these. The figures come from teardowns, regulatory filings, and developer tooling — consistent and widely corroborated, but not official.

Two splits catch people out. The base iPhone 15 has 6GB while the 15 Pro has 8GB. And iPhone 12 Pro and 13 Pro have 6GB while their non-Pro siblings have 4GB — so an older Pro device is often more capable here than a newer standard one.

The iPad Pro M4 split is by storage tier, not by model name: the 1TB and 2TB configurations ship 16GB, the smaller ones 8GB.

Putting the two tables together

12GB — iPhone 17 Pro / Pro Max, iPhone Air, 16GB iPad Pro. Everything, including the Qwen 3.5 9B flagship. On an M-series iPad Pro this is the best local AI experience available on Apple hardware, both for speed and for sustained performance without throttling.

8GB — iPhone 15 Pro, the entire iPhone 16 line, iPhone 17, and most M-series iPads. Everything except the 9B flagship. Qwen 3.5 4B is available here as the best all-around model, along with Phi-4 Mini for technical work.

6GB — iPhone 12 Pro, 13 Pro, 14 series, 15 and 15 Plus. The recommended default runs comfortably, plus Qwen 3.5 2B for better vision quality, and Llama 3.2 3B or Ministral 3 3B for general work. Qwen 3.5 4B and Phi-4 Mini are out of reach.

4GB — iPhone 12, 12 mini, 13, 13 mini. The lightweight tier: Qwen 3 0.6B and 1.7B, Llama 3.2 1B, and DeepSeek R1 1.5B. All four are genuinely useful for short questions, rewriting, summarizing, and translation, and Qwen 3 1.7B in particular is strong for its size. Multimodal models are not available at this tier.

Requirements: iOS 17.6 or later, iPhone 12 or later. The app checks your device’s memory and shows compatibility per model, so you never have to work this out yourself — but it’s useful to know what’s driving it before you buy storage you don’t need.

What affects speed but not compatibility

Catalogue estimates on an iPhone 15 Pro: Qwen 3 0.6B around 45 tokens/sec, Qwen 3.5 0.8B around 40, Qwen 3 1.7B around 30, Llama 3.2 1B around 28, Qwen 3.5 2B around 26, DeepSeek R1 1.5B around 25, Llama 3.2 3B around 20, Ministral 3 3B around 18, Phi-4 Mini and Qwen 3.5 4B around 16, and Qwen 3.5 9B around 10.

For context, comfortable reading is roughly 5–6 tokens per second, so everything in that list outruns reading on a 15 Pro. Tokens Per Second Explained covers why the numbers you actually see will be lower.

Four things move speed without changing what fits:

Chip generation. An A19 runs a given model roughly twice as fast as an A14. It cannot make a model fit that doesn’t.

Thermal design. Expect 25–40% off peak after several minutes of sustained generation as iOS manages heat. iPads and Pro models hold performance longer.

Battery state. Low Power Mode deliberately reduces peak performance.

Background apps. They compete for the same unified memory and the same bandwidth.

Practical guidance

Start with the recommended default. Qwen 3.5 0.8B is 622MB, needs 5GB, and does vision, thinking, and web search. It is the default precisely because it stays clear of the memory ceiling on heavy turns, and for most people it is enough.

Move up only when you hit a wall. The gap between the default and a larger model is smaller in daily use than the parameter counts suggest — see Is Local AI Good Enough Yet?.

Keep two models. A fast light one for everyday questions and a larger one for harder work. Switching is instant and both stay on the device.

Download over Wi-Fi, and leave storage headroom — a nearly-full device performs worse at everything.

For what each model is actually good at, see Best Local LLM Models for iPhone.

Download Cloaked on the App Store — it checks your device and shows exactly which models it can run, and everything runs entirely on-device.