Your phone can run AI.
Just not the AI you think.
No cloud, no subscription, no data leaving your pocket: modern phones run real language models fully on-device. The honest limits are the RAM an app is allowed to use and the memory bandwidth of the chip. Here's every major phone chip, what it can actually hold, and the biggest model it runs.
Phone chips, honestly compared
"Usable" is the memory an AI app can really hold — phones reserve a large share for the system, and iOS caps a single app well below total RAM. Speed estimates use the same bandwidth physics as our GPU calculator, with a conservative phone efficiency.
| Chip (phones) | Total RAM | Usable by an AI app | Bandwidth | Biggest sensible model |
|---|---|---|---|---|
| iPhone 13 (A15) iPhone | 4 GB | 2.3 GB | 34.1 GB/s | Qwen3 1.7B — comfortable |
| iPhone 13 Pro (A15) iPhone | 6 GB | 3.5 GB | 34.1 GB/s | 1.7B comfortable · 3B marginal |
| iPhone 14 Pro / 15 (A16) iPhone | 6 GB | 3.5 GB | 51.2 GB/s | 1.7B comfortable · 3B tight |
| iPhone 15 Pro (A17 Pro) iPhone | 8 GB | 5.0 GB | 51.2 GB/s | Phi-4-mini 3.8B — comfortable |
| iPhone 16 (A18) iPhone | 8 GB | 5.0 GB | 60.0 GB/s | Phi-4-mini 3.8B |
| iPhone 16 Pro (A18 Pro) iPhone | 8 GB | 5.0 GB | 60.0 GB/s | 3.8B comfortable · 7B risky |
| iPhone 17 (A19) iPhone | 8 GB | 5.0 GB | 68.2 GB/s | 3–4B comfortable |
| iPhone 17 Pro (A19 Pro) iPhone | 12 GB | 7.5 GB | 75.8 GB/s | 7–8B Q4 — plausible |
| Galaxy S23 (Snapdragon 8 Gen 2) Android | 8 GB | 5.0 GB | 67.2 GB/s | 3–4B comfortable |
| Galaxy S23 Ultra (Snapdragon 8 Gen 2) Android | 12 GB | 9.0 GB | 67.2 GB/s | 7B Q4 |
| Galaxy S24 (Snapdragon 8 Gen 3) Android | 8 GB | 5.0 GB | 76.8 GB/s | 3–4B comfortable |
| Galaxy S24 Ultra (Snapdragon 8 Gen 3) Android | 12 GB | 9.0 GB | 76.8 GB/s | 7–8B Q4 |
| Galaxy S25 (Snapdragon 8 Elite) Android | 12 GB | 9.0 GB | 84.8 GB/s | 8B Q4 — comfortable |
| Snapdragon 8 Elite · 16 GB phone Android | 16 GB | 13.0 GB | 84.8 GB/s | 14B Q4 — slow, possible |
| Dimensity 9400 · 12 GB phone Android | 12 GB | 9.0 GB | 85.3 GB/s | 8B Q4 (estimate) |
| Pixel 8 Pro (Tensor G3) Android | 12 GB | 9.0 GB | 68.2 GB/s | 3B comfortable · 7B slow |
| Pixel 9 Pro (Tensor G4) Android | 16 GB | 13.0 GB | 68.2 GB/s | 7B Q4 — modest speed |
| Pixel 10 Pro (Tensor G5) Android | 16 GB | 13.0 GB | 68.2 GB/s | 7B Q4 — modest speed |
Data reviewed Oct 2026. Bandwidth figures are the chip's theoretical LPDDR peak; real apps reach a fraction of it, which is why the calculator's phone estimates are deliberately conservative.
How do you actually run one?
PocketPal AI · iOS + Android
The default recommendation: free and open-source, built on llama.cpp, downloads models straight from Hugging Face, with a built-in benchmark. Start here.
MLC Chat · iOS + Android
Pre-compiled, GPU-accelerated models from the MLC LLM project. Where a build exists for your chip, it's usually the fastest option.
Google AI Edge Gallery · Android
Google's own offline AI showcase — the easiest 'it just works' demo on Android, built around Gemma-class models.
Russet · iOS
MLX-based and Apple-silicon only. Its developer publishes systematic on-device benchmarks — the numbers geeks trust.
Private LLM · iOS / iPad
Paid once (~$6), private by design. A favourite on iPads, where the bigger memory pool stretches to larger models.
Locally AI · iOS
A friendly chat front-end that can also use Apple's built-in on-device model — no download needed to try local AI.
The two numbers that decide everything
1 · Usable RAM = your ceiling
A model's weights have to live somewhere. A 3B model at 4-bit needs about 2.2 GB with room for conversation memory; an 8B needs about 5.5 GB. If the app can't hold it, the model doesn't run — no settings change that.
2 · Bandwidth = your speed
Generating each token means reading the whole model from memory once. Double the bandwidth, roughly double the speed. It's why two phones with the same RAM can feel completely different — and why phone AI tops out at chat speeds, not the 40+ tok/s of a desktop card.
Want more than your phone can hold?
That's the question this whole site answers. A used RTX 3060 12GB runs models no phone can touch, and when even that isn't enough, the calculator shows the cheapest card that fits — or the cloud-GPU route for the giants.
Check a computer instead →