androidai-toolstutorials

How to Use Firebase AI Logic Hybrid Inference on Android

A practical guide to Google's Firebase AI Logic Hybrid API: route Gemini requests to on-device models when they are available, then fall back to the cloud without rewriting your app.

September 17, 2026

On-device Gemini Nano is fast and private. Cloud Gemini is more capable and works on more phones. Most product features need both: a draft that still works on an airplane, and a richer answer when the network is up.

Firebase AI Logic hybrid inference is Google’s official way to use one SDK call for that split. You set a routing mode. The SDK tries the on-device model when the request is supported, then falls back to a cloud Gemini model when it is not.

This guide is for Android developers who already know how to add Firebase, and want a shipping pattern instead of a second AI stack.

What hybrid inference actually does

Firebase AI Logic can run inference on-device when a compatible model is present, and use a cloud-hosted Gemini model otherwise. Android documentation describes the same idea as a unified interface with an onDeviceConfig that picks the route.

Google’s stated reasons to prefer on-device inference:

Better privacy for the prompt and the output
Local context that never has to leave the phone
No per-call inference bill when the request stays on-device
Features that still work offline

Hybrid routing is the compatibility layer. You reach users whose phones have Gemini Nano and users whose phones do not, without two separate product paths.

On-device inference through this API is supported on Android apps using Firebase AI Logic SDK v17.10.0+ (BoM v34.10.0+) on devices that support the on-device path. It is still an experimental surface. Treat it as a preview you can ship behind a flag, not as a guarantee on every SKU in your Play Console.

What on-device can and cannot do

Official Firebase docs are strict about the on-device half:

Single-turn text generation only (not chat sessions)
Streaming or non-streaming output
Text-only input, or text plus one Bitmap image
No multi-turn memory on the device path

Cloud inference through the same SDK still supports the broader Firebase AI Logic Gemini catalog (Vertex AI in Firebase or the Gemini Developer API), including newer models such as Gemini 3.8 Flash.

That split is the design constraint. A restaurant-review draft from a photo and a few tags can stay on-device. A grounded answer that needs Maps or live web search cannot.

Google also notes that Gemini 2.5 models shut down in October 2026. New work should target a current model such as gemini-3.8-flash or gemini-3.1-flash-lite for the cloud side. Confirm the current model IDs in Firebase docs before you copy a snippet into production.

From 2 November 2026, Firebase App Check enforcement is required to use Firebase AI Logic. Plan that now if you are still on an open debug token.

The four routing modes

Android documents four InferenceMode values on OnDeviceConfig:

PREFER_ON_DEVICE — try Nano locally; if the model is missing or the request is unsupported, fall back to cloud
PREFER_IN_CLOUD — try cloud while online; fall back to on-device only when the device is offline
ONLY_ON_DEVICE — stay local; throw if on-device inference is unavailable
ONLY_IN_CLOUD — stay in the cloud; throw if the cloud path is unavailable

Start with PREFER_ON_DEVICE for private, short drafts. Use PREFER_IN_CLOUD when quality matters more than offline and you only need a last-resort local path. Use the ONLY_* modes when a silent fallback would be a product bug (for example, a compliance feature that must never leave the device).

Watch the official overview

The Android Developers talk below is a useful 45-minute pass over on-device, cloud, and hybrid work from Google I/O 2026. Watch the hybrid and Firebase sections after you read the routing table.

How to implement it

Exact artifact versions move. Follow the current Android get-started guide. The workflow does not.

1. Add Firebase AI Logic and the on-device artifact

Google’s experimental write-up uses both the core Firebase AI package and firebase-ai-ondevice. You need the on-device artifact for local routing. Pin versions from the current docs (examples in 2026 posts used firebase-ai 17.10.x and firebase-ai-ondevice 16.0.0-beta.x).

2. Create one model client with a routing mode

Android’s hybrid snippet looks like this shape:

val model = Firebase.ai(backend = GenerativeBackend.googleAI())
    .generativeModel(
        modelName = "gemini-3.5-flash",
        onDeviceConfig = OnDeviceConfig(mode = InferenceMode.PREFER_ON_DEVICE)
    )

val response = model.generateContent("Write a short review of this cafe.")

A later Android Developers Blog example used gemini-3.1-flash-lite with InferenceMode.PREFER_ON_DEVICE for a restaurant-review draft. Pick the cloud model ID from the live Firebase model list, not from an old gist.

3. Keep prompts inside the on-device contract

If you want the local path to win:

One user turn, not a chat history
One image Bitmap or none
A short instruction (“draft a two-sentence review from these tags”)
No tool use, Maps grounding, or URL context on that call

If the product later needs grounding, make a second cloud-only call. Do not hide a Maps request inside a PREFER_ON_DEVICE prompt and hope the fallback is invisible.

4. Tell the UI which path ran

Log and, when it helps the user, surface a quiet status: “Drafted on device” versus “Drafted online.” Users who care about privacy will look for that. Testers will need it when a device is on the support list but AICore has not finished downloading.

5. Handle the failure modes as product states

On-device unavailable + ONLY_ON_DEVICE → hide the button or show “not supported on this phone”
Offline + ONLY_IN_CLOUD → show a reconnect state
App Check missing after the November 2026 deadline → the call will not be a model bug
Backgrounded activity → do not assume Nano-class work is legal in a worker; ML Kit GenAI already blocks background use. Keep hybrid generation in the foreground too unless docs say otherwise

A concrete feature: review draft with a Maps link

Google’s July 2026 “cloud and hybrid inference” post walks through a restaurant feature:

The user picks topics (service, noise, food).
Hybrid inference drafts the review text, preferring on-device.
A separate cloud or Maps path can attach a place deep link when the user is online.

That is the right grain. The private sentence stays local when Nano is present. The public place identity can use the network.

Do not generate the review and the Maps deep link in one on-device call. The on-device contract does not include that grounding.

Hybrid versus ML Kit GenAI

They overlap on “run Gemini Nano locally,” but they are not the same product:

Job	Prefer
Summarize, proofread, rewrite, caption, or transcribe with a tuned task API	ML Kit GenAI
One SDK for local or cloud Gemini with explicit routing	Firebase AI Logic hybrid
Chat, grounding, or a large cloud model	Firebase AI Logic cloud path
Custom on-device prompt with no Firebase	ML Kit Prompt API or AICore preview

If the feature is “proofread this SMS,” start with ML Kit. If the feature is “draft this review on-device when we can, otherwise use Flash in the cloud,” use hybrid inference.

Practical limits to write into the spec

Experimental API: expect package names and beta versions to move
On-device path is not chat
Device support is a list, not “all Android 15+ phones”
Cloud model IDs expire; 2.5 is on a shutdown clock
App Check becomes mandatory on 2 November 2026
You still own safety, rate limits, and what you show when fallback happens

Conclusion

Firebase AI Logic hybrid inference is a routing policy, not a new model. You keep one generateContent call, set PREFER_ON_DEVICE or PREFER_IN_CLOUD, and stay inside the single-turn, optional-one-image contract when you want Nano to win.

Ship one narrow feature first: a short draft that is useful offline and acceptable when it silently uses Flash. Add App Check before November 2026. Point the cloud model at a current Gemini 3.x ID. Leave chat and grounding on the cloud path.

That is enough to stop maintaining two AI clients for the same text box.

Sources

Build hybrid and on-device experiences with Firebase AI Logic — Firebase documentation
Build hybrid experiences in Android apps — Firebase documentation
Configuration options for hybrid experiences — Firebase documentation
Hybrid inference — Android Developers
Experimental hybrid inference and new Gemini models for Android — Android Developers’ Blog (17 April 2026)
Build intelligent Android apps: Cloud and hybrid inference — Android Developers Blog (21 July 2026)
Top AI on Android updates from Google I/O 2026

Filed under: androidai-toolstutorials