For running language models on the phone itself, RAM comes first and a supported GPU second. On iOS that points to the 12 GB iPhones; on Android, to a recent Snapdragon flagship such as the Galaxy S26 Ultra. With 8 GB you can run 9B-class models, 6 GB handles 4B models well, and 4 GB is enough for small ones.
Pixels are the interesting exception: plenty of RAM, but in apps whose Android GPU path targets Qualcomm’s Adreno graphics, including Personal LLM, they run models on the CPU instead.
What matters most for running AI on a phone? #
In order:
- RAM. Decides which models fit at all. A model plus its working memory has to sit in RAM alongside the operating system. This is the one spec you can’t work around later.
- A GPU the app can use. Speeds up both reading your prompt and writing the answer, often by several times. Apps built on llama.cpp use Metal on iPhone and OpenCL on Snapdragon’s Adreno GPUs.
- Memory bandwidth. Generation speed is mostly limited by how fast the chip reads the model’s weights, so newer chips with newer memory generate faster.
- Cooling. Long sessions slow down as the phone heats up, and bigger bodies hold their speed longer.
- Storage. Each model is about 0.8 to 6 GB, plus up to 1 GB for photo support.
What matters less than the marketing suggests: NPU TOPS figures. Most local chat apps don’t use the NPU at all, which do you need an NPU to run AI explains, and what TOPS means in AI phone specs takes the number apart.
Intel, which used to feature in “AI chip wars” pieces, doesn’t make smartphone chips any more. The phone race is Apple’s A-series against Qualcomm’s Snapdragon, with Google’s Tensor, Samsung’s Exynos and MediaTek’s Dimensity behind them.
How much RAM do you need? #
Short version: 3 to 4 GB runs small models, 6 GB runs 4B-class models comfortably, 8 GB opens up 9B models, and 12 GB gives room for those plus long chats and other apps. The rule that matters when you’re comparing handsets is headroom, about 2 GB more RAM than the model’s stated minimum. How much RAM you need to run an LLM on a phone has the full table of model sizes against RAM tiers.
If you plan to load your own larger models later, buy more RAM than you need today.
What are the best iPhones for local AI? #
Every iPhone uses its GPU through Metal, so RAM is the only variable that changes the answer. Apple doesn’t publish RAM figures; these come from teardowns and reviews.
| iPhone | RAM | What it runs well |
|---|---|---|
| iPhone 17 Pro, 17 Pro Max, iPhone Air | 12 GB | Everything in a typical catalog, including 6 GB vision models |
| iPhone 17, iPhone 16 family (incl. 16e), iPhone 15 Pro and Pro Max | 8 GB | 9B models comfortably |
| iPhone 15, 15 Plus, 14 family, 13 Pro and Pro Max | 6 GB | 4B models comfortably; 9B at a squeeze |
| iPhone 13, 13 mini | 4 GB | Small and 4B-class models |
The base iPhone 17 is the value pick: 8 GB runs the strongest general models most catalogs offer. The 12 GB phones buy headroom for longer chats, image support and larger custom models.
Local AI apps typically need only iOS 15.1 or later, so iPhones far too old for Apple Intelligence still run open models perfectly well. See does Apple Intelligence work offline for how the two differ.
What are the best Android phones for local AI? #
On Android the chip matters twice: for how much RAM the phone ships with, and for whether the app can use the GPU at all. GPU acceleration in Personal LLM uses OpenCL on Snapdragon phones with Adreno 700-series or newer GPUs, which in practice means flagship Snapdragon 8-series chips from the 8 Gen 1 onward. Other Android phones run on the CPU, which works but is slower.
| Phone | RAM | Chip and GPU | GPU acceleration via Adreno |
|---|---|---|---|
| Galaxy S26 Ultra | 12 GB (16 GB with 1 TB storage) | Snapdragon 8 Elite Gen 5, Adreno 840, worldwide | Yes |
| Galaxy S26, S26+ | 12 GB | Snapdragon in North America, China and Japan; Exynos elsewhere | Snapdragon models only |
| Galaxy S25 series | 12 GB | Snapdragon 8 Elite, Adreno 830 | Yes |
| Other recent Snapdragon 8 Elite flagships | Often 12 to 16 GB | Adreno 830 or 840 | Yes |
| Pixel 10 Pro, 10 Pro XL | 16 GB | Tensor G5, non-Adreno GPU | No, runs on CPU |
| Pixel 10 | 12 GB | Tensor G5, non-Adreno GPU | No, runs on CPU |
The Galaxy S26 Ultra is the safest Android pick because it ships with Snapdragon everywhere. If you’re buying an S26 or S26+ outside North America, China or Japan, check which chip your regional model uses before counting on GPU acceleration; an Exynos Galaxy falls back to the CPU.
Pixels are a genuine mixed case. Their RAM is generous, so a Pixel 10 Pro can hold a large model without trouble, but it generates on the CPU in apps that accelerate only Adreno GPUs. Run a benchmark and decide whether the speed suits you rather than assuming either way.
What are the best budget and used phones for local AI? #
You don’t need this year’s flagship. Older Snapdragon flagships keep their GPU support, and 8 GB still runs a 9B model.
- Galaxy S24 Ultra (12 GB, Snapdragon 8 Gen 3, Adreno 750). The strongest used Android pick, and every S24 Ultra uses Snapdragon.
- Galaxy S23 series (8 GB on most models, Snapdragon 8 Gen 2, Adreno 740). Runs 9B models with GPU acceleration.
- Galaxy S24 and S24+ use Snapdragon in some regions and Exynos in others, so check the exact model.
- iPhone 15 Pro (8 GB). Runs the same models as a current 8 GB iPhone.
- iPhone 14 or 13 Pro (6 GB). Fine for 4B-class models, which cover most everyday use.
A used flagship with 8 GB beats a new budget phone with 4 GB for this purpose, every time.
How much storage do models take? #
Between about 0.8 and 6 GB per model, plus 195 MB to 1 GB if you add image support. Two or three models fit easily on a 128 GB phone. How much storage AI models need has the per-model numbers and a budget.
How do you check the phone you already have? #
Before buying anything, test what’s in your pocket:
- Install a local AI app and open the model catalog. In Personal LLM each model shows a “Fits your device” or “Should run” badge based on your actual RAM.
- Download a 4B model, the sensible starting point on most phones.
- Run the built-in benchmark. It reports load time, generation speed in tokens per second, and whether the GPU or the CPU did the work. That last line answers the Adreno question for your specific phone without any spec-sheet research.
- Try a real task and watch the live tokens-per-second readout. If replies appear faster than you read, your phone is already good enough.
Local LLM speed on your phone explains how to read those numbers and what counts as fast.
Frequently asked questions #
What’s the best phone for running AI locally? #
A 12 GB iPhone (17 Pro, 17 Pro Max or iPhone Air) or a recent Snapdragon flagship such as the Galaxy S26 Ultra. Both run every model in a typical phone catalog with GPU acceleration. The base iPhone 17 and recent 8 GB Snapdragon phones are strong, cheaper alternatives.
How much RAM do I need to run an LLM on my phone? #
About 3 to 4 GB for small models, 6 GB for 4B-class models, and 8 GB or more for 9B models. More RAM also leaves room for longer conversations and image support.
Can a Pixel run local AI models? #
Yes. Recent Pixels have 12 to 16 GB of RAM, enough for large models. In apps whose Android GPU acceleration targets Snapdragon’s Adreno GPUs, a Pixel runs on its CPU, so check a benchmark to see whether the speed works for you.
Is iPhone or Android better for local LLMs? #
Neither wins outright. Every iPhone gets GPU acceleration through Metal, so choosing one is only about RAM. On Android, a Snapdragon 8-series flagship gets GPU acceleration and often more RAM for the money, while other chips fall back to the CPU. iPhone vs Android for local AI compares them in detail.
Does a gaming phone run AI better? #
Sometimes, and not for the reason you’d expect. Gaming phones often pair a Snapdragon flagship with lots of RAM and better cooling, which helps sustained speed. The gaming-specific features do nothing for a language model.