Google Gemma 4 Supports Fully Offline AI Inference on iPhone

Google's open-source model Gemma 4 can now run natively on iPhone without an internet connection, performing complete offline inference.

Model Performance and Optimization

• The Gemma 4 31B model shows performance similar to the Qwen 3.5 27B model.
• E2B and E4B variant models optimized for mobile environments are provided. In particular, E2B is recommended by default in the Google app due to its advantages in speed, lightweight footprint, and thermal control.

Execution and Usage

• Users can select models and run inference directly on-device through the Google AI Edge Gallery app on the App Store, without any API or cloud dependency.
• The app goes beyond a simple text interface, serving as an on-device AI experimentation platform that includes image recognition, voice interaction, and an extensible Skills framework.

Technical Features

• By utilizing the iPhone's GPU during inference, it achieves low latency, demonstrating that high-performance AI workloads can be handled even on consumer hardware.
• However, since it uses the GPU rather than Apple's dedicated accelerator, the ANE (Apple Neural Engine), there remains room for future optimization in terms of battery efficiency.