Google's Gemma 4 open-source model family now executes directly on iPhones with complete local inference and offline capability. Users download Google AI Edge Gallery from the App Store, select a model variant, and begin running inference without cloud dependency or API calls.
Gemma 4 arrives in three deployment tiers. The 31B variant benchmarks near Qwen 3.5's 27B model, with Gemma carrying approximately 4 billion additional parameters. Both models carry trade-offs across different tasks with no clear winner across every benchmark.
The more significant story lies in the E2B and E4B variants, engineered specifically for mobile deployment. These models prioritize efficiency over raw capability, optimized for the thermal and memory constraints of consumer hardware. Google's own AI Edge Gallery defaults users toward E2B, reflecting the practical reality that speed and efficiency matter more than parameter count in on-device scenarios.
Google AI Edge Gallery functions as more than a text interface. The platform bundles image recognition, voice interaction, and an extensible Skills framework, positioning itself as a foundation for on-device AI experimentation rather than a limited demo or feature.
Inference routes through the iPhone's GPU. Responses arrive with notably low latency, a concrete indicator that consumer hardware can sustain this class of workload without degradation. This latency profile changes the commercial viability calculation for local AI deployment.
Offline capability carries particular weight for enterprise applications. Field operations, healthcare settings, and environments where data privacy regulations prohibit cloud processing now have a viable option for deploying capable AI models on device.
The shift from theoretical edge AI to functional deployment on mainstream consumer hardware marks a meaningful inflection point. Gemma 4 on iPhone demonstrates that on-device AI capability has moved from future prospect to present reality.