Google’s open model Gemma 4, released on April 2, can now run full local inference on iPhone. Users can download Google AI Edge Gallery from the App Store, pick a model variant, and start using it without a network connection or any cloud processing.
This rollout goes beyond a limited demo. Google has put offline inference directly into a consumer-facing app, removing the usual setup barriers such as developer accounts and API keys.
Google moved complete model inference onto the device
Gemma 4 is released under the Apache 2.0 license and comes in four open-weight variants. On iPhone, the inference path uses the GPU rather than Apple’s Neural Engine. The rationale is technical: Neural Engine is optimized for fixed-pattern workloads such as Face ID-style recognition, while large language model inference needs more flexible floating-point computation, making the GPU the more practical route.
Google AI Edge Gallery is not limited to a text interface. The app also combines image recognition, voice interaction, and an extensible Skills framework, giving it the shape of an experimental development platform rather than a single-purpose showcase.
E2B is the default path for mobile use
On benchmark comparisons, the 31B flagship version of Gemma 4 posts results close to the 27B version of Qwen 3.5. Each model has strengths on different tasks, and the source material makes clear that this head-to-head comparison is not the main point of the deployment strategy.
Google’s app steers users toward the E2B variant by default for practical reasons: lower memory use, lower thermal load, and lower power draw. E2B and E4B were designed with offline mobile environments in mind, prioritizing sustained usability over benchmark-focused performance.
The suggested hardware floor is an iPhone 15 Pro or newer. That places the current target audience firmly in the early-adopter group for on-device AI.
Offline inference matters most where data cannot leave the device
The real change is not just lower latency. Local inference expands what can be done in settings where sensitive data must stay on site, such as patient records in medical institutions, image logs from field inspections, or confidential internal company files. In those cases, the issue with cloud AI is not primarily speed; it is compliance.
Seen from that angle, Gemma 4 on iPhone offers a way around data-transfer restrictions rather than a direct performance challenge to cloud-based systems. The source also notes that many current use cases remain in the proof-of-concept stage, with security audits, management tooling, and integration costs still standing between basic functionality and broad deployment.