Google Gemma 4 Runs Natively on iPhone, Supports Fully Offline AI Inference

Google's Gemma 4 runs directly on iPhone, performing local inference without an internet connection.

The key point is not just a demo — it signals that on-device AI on mobile devices has entered the actual deployment stage. Google provides this model through the Google AI Edge Gallery app, and users can download the app from the App Store, then select the desired model variant to run it immediately.

Model selection emphasizes device suitability over raw performance.

• The 31B variant is mentioned as being in a similar class to Qwen 3.5 27B.
• E2B and E4B are small models designed with mobile deployment in mind, prioritizing efficiency and lightweight footprint.
• The Google app specifically recommends E2B, explaining that it is better suited to real-world environments with memory and heat constraints.

The functionality goes beyond a simple text chatbot. It includes image recognition, voice interaction, and an extensible Skills framework, making it closer to a platform for on-device AI experimentation.

Inference runs through the iPhone's GPU, with low response latency, sending the message that this kind of workload is now fully feasible on consumer hardware.

Offline operation is especially meaningful in environments such as:

• field applications
• healthcare settings
• work where data transfer is restricted or privacy is critical

Ultimately, the point of this announcement is that local AI on mobile devices is no longer a distant concept. Through Gemma, Google is pushing a practical inference foundation for edge devices, including iPhone.