Gemma 4 goes fully offline on mobile, no cloud required

Google has made Gemma 4 available on-device through React Native, enabling cross-platform mobile apps to run the model entirely offline with local hardware acceleration. The integration supports the Vulkan delegate on Android devices and the MLX delegate on Apple Silicon, tapping into GPU and neural engine capabilities for inference. Local LLMs just ate cloud triage for lunch

This deployment unlocks Gemma 4's vision and tool-use features without requiring a server round trip. Developers can build apps that read documents, analyze images, or trigger actions such as scheduling events, all processed on the user's device. In a demo, the model reads a printed flyer and adds a calendar entry, completing the workflow with no connectivity. IBM's new open-source agent framework cuts the…

Local execution addresses latency and privacy concerns by keeping data on-device. It also reduces reliance on cloud infrastructure, cutting operational costs for developers. React Native's cross-platform nature means a single codebase can support both Android and iOS deployments with minimal adaptation. Semi-autoregressive decoding just broke the 85% speed…

Performance benchmarks for the on-device Gemma 4 variant were not shared, but earlier versions of the model family have shown competitive results for mobile-sized language models. The tool-use capability, which allows the model to call functions and interact with app components, is a differentiator for on-device AI assistants that need to perform real tasks. Microsoft's new platform gives scientists a governed…

Developers can start testing the integration today through Google's updated ML Kit and MediaPipe packages for React Native. Documentation and sample code are available for those building next-generation mobile experiences with local AI. China's tech giants just deleted millions of AI…