Gemini Nano vs Cloud-Based Gemini Models

Our Cloud vs On-Device AI article compares the approaches in general. This one is specific to Google's Gemini family on Android: what Nano can and can't do, what cloud models add, and how to use both.

Two paths to Gemini

Side-by-side

Model names and limits change frequently. Current cloud models include gemini-3.8-flash and gemini-3.5-flash-lite. Check the models page before shipping.

Hybrid inference

Firebase AI Logic offers experimental hybrid inference: the SDK uses on-device Gemini Nano when available, otherwise the cloud model:

val model = Firebase.ai(backend = GenerativeBackend.googleAI()).generativeModel(
    modelName = "gemini-3.8-flash",                       // cloud fallback
    onDeviceConfig = OnDeviceConfig(mode = InferenceMode.PREFER_ON_DEVICE),
)
val reply = model.generateContent("Rewrite politely: $draft").text

Because the experimental API may change, follow the hybrid inference docs for the exact dependency and names.

How to choose

1. Sensitive data, offline need, simple task → Nano where available, with a non-AI fallback elsewhere.
2. Must work on every device, harder task → cloud Gemini.
3. Simple task, wide reach, cost-sensitive → hybrid: Nano when present, cloud otherwise.
4. Long documents or complex reasoning → cloud.

Use cases

Frequently asked questions

Conclusion

Gemini Nano offers private, offline, free inference for simple tasks on supported devices. Cloud Gemini offers stronger models everywhere at a per-token cost. Hybrid inference combines them, with clear disclosure.

Next step: run your own models on device with LiteRT and LiteRT-LM.