Gemini Nano on Android
by Google DeepMind
Google’s on-device AI model for Android — delivered through AI Core for privacy-preserving, offline-capable inference on mobile devices.
See https://ai.google.dev/edge/mediapipe/solutions/genai/llm_inference
Features
- On-device inference via AI Core: centralized model delivery, hardware optimization, scheduling, and privacy isolation
- ML Kit GenAI APIs: summarization, proofreading, rewrite, and a prompt API for natural-language requests
- Prompt API supports text and image input with text output
- Firebase AI Logic for hybrid inference: use Gemini Nano on supported devices, fall back to cloud (Gemini API, Vertex AI, Gemini Developer API) when unavailable
- Cloud fallback models: Gemini Flash, Flash Lite, and Pro
- RAG-like workflows supported via the prompt API; embedding API coming soon
- Battery and RAM managed at the system level by AI Core; foreground apps prioritized
- Classical ML Kit models (OCR, vision) run on many devices; GenAI currently targets newer flagship-class devices
- LightRAG/Light RT LM path available for custom-model deployment on broader device range
Superpowers
Gemini Nano’s on-device deployment through AI Core is a fundamentally different approach to mobile AI: instead of every app shipping its own model, the system manages one shared, optimized model for all apps. This means developers get privacy, offline capability, and low latency without managing model downloads, weights, or hardware optimization. The hybrid inference pattern via Firebase AI Logic is particularly practical — detect whether the device supports Gemini Nano, use it when available, fall back to cloud otherwise. For sensitive use cases (personal data, health, finance), on-device inference keeps data local entirely. The emerging embedding API and RAG prompt support point toward increasingly capable on-device AI applications.
Pricing
Free for developers via ML Kit and Firebase AI Logic. Cloud fallback uses Gemini API pricing.