Google’s AI Edge Gallery app enables users to run powerful, offline Gemma-based large language models directly on mobile devices to perform assistant tasks and local data processing.

Key Points

• The app utilizes Google’s Gemma 4 E2B and E4B models, which support multimodal inputs and context windows of up to 32k tokens.
• Agent Skills allow for offline calendar management, notification scheduling, and image identification without requiring an internet connection.
• FunctionGemma 270M powers specific mobile actions, such as controlling device hardware like the flashlight or volume via natural language commands.
• Users can monitor step-by-step reasoning through a "Thinking Mode" and perform device-specific performance benchmarking for tokens-per-second metrics.
• The platform supports sideloading custom .litertlm files and Hugging Face imports for users who wish to run models outside the standard Gemma catalog.

Why it Matters

This development signals a significant shift toward privacy-focused, offline AI utility that reduces reliance on cloud-based processing for everyday smartphone tasks. By bringing sophisticated model performance to mobile hardware, Google is lowering the barrier for users to integrate local LLMs into their daily workflows.