Google Shifts AI From the Cloud to Devices

For years, artificial intelligence has largely lived in the cloud, with large models running in centralized data centers powering chatbots, enterprise tools, and consumer apps. While effective, this approach comes with limits, including delays caused by network connections, rising infrastructure costs, and the need to move user data across networks. As AI becomes part of operating systems and everyday software, these trade-offs are becoming harder to ignore.

Google is now signaling a shift in how it wants AI to work. Alongside its cloud-based Gemini models, the company is expanding its edge AI strategy, focusing on running models directly on devices. A key part of this push is FunctionGemma, a compact model designed to operate on mobile hardware without relying on cloud processing. This reflects Google’s view that local execution should be a core layer of AI infrastructure, not just an optimization.

FunctionGemma is built to turn natural language commands into direct actions on a device. Instead of generating conversational text, it produces structured instructions that software can execute. This allows phones and other devices to respond instantly to user intent, even when internet access is limited or unavailable. Because processing happens locally, actions occur without delays caused by network round-trip times, and sensitive data stays on the device rather than being sent to external servers.

The model is based on a smaller version of Google’s Gemma family but is trained specifically for reliability in executing actions. Internal testing showed that general-purpose small models often struggle with precise tasks, while FunctionGemma’s targeted training significantly improved accuracy. This specialization highlights a broader shift in how AI is designed, focusing on doing tasks reliably rather than simply explaining them.

FunctionGemma’s small size is central to its role. It is designed to run on constrained hardware while maintaining enough context to handle real-world commands. Google does not position it as a standalone assistant but as an embedded component that quietly enables action-focused AI within apps and systems.