Gemini Nano for Android developers

The feature-specific APIs are already tuned: low integration effort, no prompt engineering, but fixed behavior, such as summaries of 1 to 3 bullet points, generic short image descriptions and predefined rewriting styles. The Prompt API accepts text or image plus text, returns text or structured output, and suits tasks like image classification, short translations, guided summaries, entity extraction and receipt scanning, at the cost of more prompt engineering and quality testing. The feature APIs also run on a somewhat different device list, including the Galaxy S25 series. Prompt API is in beta, with no SLA or deprecation policy, and can change in breaking ways.

The Prompt API needs Android API level 26 or higher; Google’s getting-started page adds it as com.google.mlkit:genai-prompt:1.0.0-beta4. Get a client with Generation.getClient(), then call checkStatus(): UNAVAILABLE means the device isn’t supported or hasn’t fetched its configuration yet, DOWNLOADABLE means you can call download(), and AVAILABLE means the model is ready. Because apps share the one Gemini Nano model that AICore manages, it doesn’t count against your app’s size or memory budget. Use getBaseModelName() to see which Nano version a device runs, since versions can answer the same prompt differently. Optional parameters include temperature, topK and candidateCount, and you can stream responses.

AICore enforces an inference quota per app: too many requests in a short time return ErrorCode.BUSY, so retry with exponential backoff, and exceeding a longer quota, such as a daily one, returns PER_APP_BATTERY_USE_QUOTA_EXCEEDED. Inference only runs while your app is the top foreground app; calls from the background, including from a foreground service, return BACKGROUND_USE_BLOCKED. Google doesn’t publish the quota numbers.

AICore follows Android’s Private Compute Core principles: it is isolated from most other packages, has no direct internet access, with model downloads routed through the open-source Private Compute Services app, and doesn’t keep a record of inputs or outputs after processing. Apple’s equivalent is covered in Apple Foundation Models; to expose your own features to Gemini, see Android AppFunctions.

An on-device AI feature gives people a reason to install; being found is what gets them to your listing. Censuus ranks apps by the visits they draw, and adding your app is free on the List my app form. Placement comes from traffic and sponsorship: apps climb on the visits they draw, and a sponsor can pay to rise higher.