Run Gemma 4 on iPhone 15 Pro with Google AI Edge Gallery

iPhone 15 Pro was the first 8 GB iPhone, and that extra memory is exactly what makes it a viable target for Gemma 4 in 2026. The A17 Pro chip handles Gemma 4 E2B at an estimated 22 tokens per second through Google AI Edge Gallery, slower than newer iPhones but still fast enough for comfortable conversation. If you already own this phone, you do not need to upgrade for serious on-device AI.

Apple A17 Pro8 GB RAM2023

Quick answer

On the iPhone 15 Pro (Apple A17 Pro, 8GB RAM), run Google Gemma 4 E2B at ~22 tok/s (est.) via Google AI Edge Gallery. Gemma 4 E2B is ~2.5 GB on disk and uses 1-1.5 GB in memory, fully offline, no account required.

Sizing rule: an on-device model must fit its weights plus working memory inside the iPhone 15 Pro's 8GB of RAM, shared with iOS and the host app. The larger Gemma 4 E4B (~5GB on disk, 2-3GB in use) is the quality upgrade, but it only fits phones with more free memory. Everything runs fully offline once the model is downloaded, and speed varies with thermal state during long sessions. What it will not run: 7B-class and larger models exceed a phone's usable memory; those need a Mac, a GPU, or a cloud API.

→ Install Google AI Edge Gallery (iOS 17+), then download Gemma 4.

TOP PICK

Gemma 4 E2B

SPEED

~22 tok/s (est.)

DEVICE RAM

8 GB

Speed is a ModelFit estimate from chip generation, not a measured benchmark.

Cite this page: ModelFit, Run Gemma 4 on iPhone 15 Pro: Edge Gallery Guide (2026), https://modelfit.io/iphone/iphone-15-pro/, updated August 2026, CC BY 4.0.

Last updated: August 26, 2026 · Editor: ModelFit Team

Verdict

Recommended: Gemma 4 E2B

iPhone 15 Pro is the oldest iPhone where Gemma 4 feels genuinely fast. 8 GB RAM and A17 Pro keep it relevant.

Hardware Profile

DEVICE

iPhone 15 Pro

CHIP

Apple A17 Pro

RAM

8 GB

NEURAL ENGINE

16-core

Gemma 4 Performance on iPhone 15 Pro

Bar chart: Gemma 4 speed on iPhone 15 Pro in Google AI Edge Gallery. Gemma 4 E2B ~22 tok/s, Gemma 4 E4B ~11 tok/s.

Model	Download	RAM in Use	Speed	Source
Gemma 4 E2BTop Pick	~2.5 GB	1-1.5 GB	~22 tok/s	Estimated
Gemma 4 E4B	~5 GB	2-3 GB	~11 tok/s	Estimated

Speeds via Google AI Edge Gallery on iOS 17+. "Measured" numbers come from real-world Hacker News user reports; "Estimated" numbers are interpolations from chip generation. Both Gemma 4 variants use int4 quantization-aware training.

Best for

iPhone 15 Pro and 15 Pro Max owners who want to stay current
Gemma 4 E2B for daily chat and writing
Offline use during travel or low-signal areas
No-upgrade path to capable on-device AI

Watch outs

A17 Pro is roughly 30% slower than A18 Pro at neural workloads
Gemma 4 E4B is borderline. Prefer E2B
Sustained inference will warm the device under heavy use

Setup Guide

Step-by-step install for Google AI Edge Gallery on iPhone 15 Pro, plus full benchmarks and the privacy details.

Read the full guide

Frequently Asked Questions

Does iPhone 15 Pro support Google AI Edge Gallery?

Yes. iPhone 15 Pro runs iOS 17 and later and meets the Google AI Edge Gallery system requirements. Both Gemma 4 E2B and E4B variants are supported, with E2B as the recommended daily-use pick.

How fast is Gemma 4 on iPhone 15 Pro?

Estimated at roughly 22 tokens per second for Gemma 4 E2B on the A17 Pro. That is slower than the iPhone 16 Pro's measured 30 tok/s but still fast enough for fluent, conversational use. We have not seen measured real-world numbers yet for this exact device.

Is iPhone 15 Pro still good for local AI in 2026?

Yes. With 8 GB of RAM and the A17 Pro's 16-core Neural Engine, iPhone 15 Pro handles Gemma 4 E2B and Apple Intelligence comfortably. The newer iPhone 16 Pro and 17 Pro are faster, but the gap is incremental. There is no urgent reason to upgrade for AI alone.

Can iPhone 15 Pro run Gemma 4 E4B?

Technically yes, with caveats. The 5 GB download fits in storage, and the 2-3 GB active memory load works on 8 GB RAM, but iOS will aggressively close background apps. Expect ~11 tok/s. For sustained E4B use, iPhone 17 Pro with 12 GB RAM is the better target.