Gemini Nano Explained: AICore, ML Kit GenAI APIs, Chrome and Supported Devices

Gemini Nano is the model behind a growing number of AI features that run on phones, laptops and inside Chrome without sending prompts to a server, and its rules differ sharply from cloud Gemini. Gemini Nano is Google's on-device model in the Gemini family, first released on the Pixel 8 Pro on 6 December 2023, which Google describes as its most efficient model for on-device tasks. On Android it runs inside the AICore system service and is reached through the ML Kit GenAI APIs; in desktop Chrome it powers the built-in AI APIs such as the Prompt, Summarizer and Translator APIs. This page sets out what Gemini Nano is, which versions exist, how developers and users reach it, what hardware it needs, what Google says about data handling, and the limits that matter in practice. It is written for developers, product owners, mobile and browser administrators and privacy teams who want facts from Google's own documentation.

Last checked: 2026-11-15.

This is an independent reference page from NHI Mgmt Group based on public provider documentation; it is not sponsored or endorsed by any provider.

Key takeaways

• Two runtimes, one model family. On Android, Gemini Nano runs in the AICore system service and is reached through ML Kit GenAI APIs; on desktop Chrome it sits behind the built-in AI web APIs.
• Three model versions in the device tables. Google's ML Kit device tables list Gemini Nano 2, 3 and 4, and apps can read the installed version with getBaseModelName().
• No per-call price. Inference runs on the user's hardware, so Google lists no token price for Gemini Nano; cloud fallback through Firebase AI Logic is billed as cloud Gemini.
• Local by design, with stated limits. Google says AICore does not store a record of inputs or outputs after processing and has no direct internet access, but inference is foreground-only and subject to per-app quotas.
• Chrome needs real hardware. Chrome's foundation-model APIs need a desktop with at least 22 GB free storage and either a GPU with more than 4 GB of VRAM or 16 GB of RAM with 4 CPU cores, and do not run on Chrome for Android or iOS.

What is Gemini Nano?

Google announced Gemini 1.0 on 6 December 2023 in three sizes, Ultra, Pro and Nano, and named Nano "our most efficient model for on-device tasks". Android's launch post the same day described Nano as distilled from the larger Gemini models and optimised for mobile silicon accelerators, with summarisation, contextual smart replies and proofreading named as early capabilities. The Pixel 8 Pro was described as the first smartphone engineered to run it.

The model itself is not downloaded by app developers. On Android it is a system component that the operating system downloads, updates and runs; in Chrome the browser downloads and manages it. Developers call an API, and the platform decides whether the model is available on that device. Google does not publish a parameter count for current Gemini Nano versions on the pages checked for this guide.

Key facts

Which Gemini Nano versions exist?

The ML Kit GenAI overview groups supported devices by base model version. Google notes that availability can vary by RAM configuration, OS version and AICore update status, and that different versions may return different output for the same prompt.

As of 2026-11-15, from the ML Kit GenAI overview (last updated 10 November 2026). Device lists change often; check the live table before planning a release. The Chrome documentation does not label the browser model with these version numbers and points developers to chrome://on-device-internals for the current model size.

How do developers use Gemini Nano on Android?

Apps do not talk to the model directly. They call ML Kit GenAI APIs, which run on top of AICore. Google's overview labels each surface with a maturity stage.

As of 2026-11-15. No ML Kit GenAI surface carried a general availability label on the overview checked. Use of the APIs is governed by the ML Kit GenAI API Additional Terms of Service, and Google states developers are solely responsible for the safety of their API client and their app's user experience.

• Version check: call getBaseModelName() to see which Gemini Nano version is installed, and test prompts against each version you intend to support.
• Hybrid fallback: the Firebase AI Logic Hybrid API can route a request to Gemini Flash or Gemini Pro in the cloud when Nano is unavailable or unsupported, and protect the cloud path with Firebase App Check and attestation such as Play Integrity.
• Customisation history: the December 2023 AICore launch post described LoRA adapters that developers could train on their own data for AICore to load. The current overview does not describe this as a public feature, so treat it as historical.
• Early access: Google runs an AICore Developer Preview programme for features ahead of wider release.

How does Gemini Nano work in Chrome?

Chrome's built-in AI lets a website or extension call models that the browser provides and manages. The first call to an API triggers the model download; after that the model runs locally with no network connection needed.

APIs listed alphabetically. Status as of 2026-11-15 from Chrome's built-in API status page; a separate web origin trial covers Prompt API sampling parameters from Chrome 148. Some APIs are limited to Early Preview Program participants, and an extension that uses an API in origin trial must register for that trial.

• Availability check: availability() returns unavailable, downloadable, downloading or available. If the model is not yet downloaded, the user must interact with the page (a click or key press) before create() can start a session.
• Context handling: the Prompt API exposes session.contextWindow and session.contextUsage. On overflow the oldest prompt and response pairs are dropped while the system prompt is kept, and a contextoverflow event fires; a prompt that cannot fit fails with QuotaExceededError.
• Structured output: a JSON Schema or regular expression passed as responseConstraint shapes the response, and counts against the context window unless omitResponseConstraintInput is used.
• Sampling: web pages use a samplingMode option with seven values from "most-predictable" to "most-creative"; extensions set topK and temperature together or not at all.
• Hybrid option: Chrome's documentation also points to Firebase AI Logic for a cloud fallback when the device cannot run the model.

For extension permissions and review, see the Chrome Manifest V3 Transition.

Which devices support Gemini Nano?

As of 2026-11-15. Language Detector and Translator have no OS or storage requirement listed beyond desktop Chrome. For managed fleets see the ChromeOS and Chromebook Security Guide and the Android Enterprise Security Guide.

How much does Gemini Nano cost?

The cost to watch is not money but device resources: Google's ML Kit documentation returns ErrorCode.PER_APP_BATTERY_USE_QUOTA_EXCEEDED when an app passes a long-duration quota such as a daily limit.

What features does Gemini Nano power?

Google's own products were the first users of the model. The features below are those named on official pages checked for this guide.

In the Chrome scam feature, only users who opted into Enhanced Protection have the model-generated signals sent to Safe Browsing; Standard Protection users benefit when newly found sites are added to blocklists. Google said in May 2025 it planned to bring the feature to Chrome on Android later that year.

Provenance and labelling

Gemini Nano's developer surfaces mostly produce text, image captions, embeddings and speech. Google's ML Kit and Chrome pages checked for this guide do not state that outputs carry a SynthID watermark or C2PA Content Credentials: not stated by Google. Apps that generate or edit media with other models should follow their own labelling duties. For background on the two main approaches see SynthID vs C2PA Content Credentials and the AI Content Provenance and C2PA Guide.

Data and privacy

• No stored records: Google states AICore isolates each request and does not store any record of the input data or the resulting outputs after processing.
• Network isolation: AICore has no direct internet access. All internet requests, including model downloads, go through the open-source Private Compute Services companion APK.
• Restricted binding: AICore is isolated from most other packages, with limited exceptions for specific system packages, and that allow list changes only through a full Android OTA update.
• Private Compute Core: AICore follows Android's Private Compute Core principles.
• Safety filters: AICore evaluates requests against Google's safety filters.
• Chrome: after download the model runs locally; Google describes client-side AI as offering increased user privacy. Where a developer adds a hybrid cloud path, data sent to the cloud follows the cloud provider's terms.

On-device does not mean risk-free: the app holding the data, the prompt it builds and anything it sends to a cloud fallback still need review. The On-Device AI Security Guide covers threat models for local models.

When was Gemini Nano released, and how has it changed?

Chrome version rows are dated by release, not calendar date, because Google's API pages state the version only.

What are Gemini Nano's limits?

• Foreground only on Android: inference is allowed only when the app is the top foreground application; background use, including from a foreground service, returns ErrorCode.BACKGROUND_USE_BLOCKED.
• Rate and quota limits: ErrorCode.BUSY signals requests arriving too quickly (Google recommends exponential backoff); long-duration quotas return PER_APP_BATTERY_USE_QUOTA_EXCEEDED.
• Capability bounds: Google lists token limits, device hardware limits and model capability boundaries as the main constraints, and recommends hybrid fallback for harder tasks.
• Version drift: the same prompt can give different output on Nano 2, 3 and 4.
• Pre-GA surfaces: every ML Kit GenAI surface checked is Alpha, Beta, Preview or Experimental, and three Chrome APIs are developer trials.
• Languages: Chrome's models support five languages; on Android, language support depends on device configuration and downloaded models, and Google publishes no single list.
• Storage churn: Chrome purges the model below 10 GB free space, so availability can change on the same machine.

Where identity and security fit

Hybrid designs move the identity question to the cloud path: Google recommends Firebase App Check with attestation such as Play Integrity, and keeping sensitive system prompts server-side, so that only genuine app instances reach the paid model. Treat any API key used for the fallback as a secret with an owner and rotation plan. See the Android App Security Guide.

Practitioner checklist

• Check the live ML Kit device tables and record which Gemini Nano versions your supported devices run.
• Call getBaseModelName() in testing and run your prompt suite against each Nano version.
• Design for BACKGROUND_USE_BLOCKED, BUSY and battery quota errors, with backoff and a user-visible fallback.
• If you add a cloud fallback, protect it with Firebase App Check and keep system prompts server-side.
• In Chrome, call availability() first, explain the model download to users and require a user gesture before create().
• Do not ship features that depend on developer-trial Chrome APIs to general users without a fallback.
• Tell privacy teams which data stays on device and which leaves through any hybrid path, and update your privacy notice.
• For managed desktops, confirm the 22 GB storage and GPU or RAM requirements before promising Chrome AI features.

Standards and references

• Android Developers: Gemini Nano
• Google for Developers: ML Kit GenAI APIs
• Chrome for Developers: Built-in AI APIs status
• Chrome for Developers: Get started with built-in AI
• Chrome for Developers: The Prompt API
• Android Developers Blog: A new foundation for AI on Android
• Google: Introducing Gemini
• Google: Using AI to stop tech support scams in Chrome

Related NHI Mgmt Group resources: On-Device AI Security Guide · What Is an NPU? TOPS, AI PC Chips and What the Numbers Mean · Android Enterprise Security Guide · SynthID vs C2PA Content Credentials