Gemini Intelligence's 12GB RAM floor: an Android AI fallback plan for 2026

Summary. Google's requirement list for Gemini Intelligence sits in footnote 1 of android.com, and it names eight conditions including "12GB+ RAM", a "Qualified SOC (flagship chip)" and "Nano v3 or greater". Gemini Nano itself runs on more than 140 million devices as of 21 July 2026, and the latest version, Gemini Nano 4, is built on the Gemma 4 architecture. Those two numbers describe different populations, and confusing them is the most common planning error we see. The ML Kit GenAI APIs give you six on-device features, all still labelled Beta except Speech recognition at Alpha, and they require Android API level 26 or higher. Google's own answer for devices that miss the bar is the Firebase AI Logic Hybrid API, with a PREFER_ON_DEVICE routing mode that falls back to a cloud model automatically. That fallback has a price: Gemini 3.1 Flash-Lite costs $0.25 per 1M input tokens and $1.50 per 1M output tokens on the paid standard tier, or $0.125 and $0.75 through the Batch API. Google does not publish RAM distribution across the Android installed base, so the only way to size your own gap is the Reach and devices report in Play Console.

This is a build guide, not a launch recap. It answers one question: given that a hard hardware floor exists and most of your installed base sits under it, what does a shippable Android AI feature actually look like in 2026.

What Google actually requires, in Google's words

Most coverage of Gemini Intelligence device requirements traces back to trade reporting. The primary source is narrower and stricter than the summaries. Google's Gemini Intelligence page carries this footnote:

"Gemini Intelligence features are only available on Android devices with the most advanced capabilities and spec requirements: (1) Nano models on device AI (integrate AI Core + Nano v3 or greater) (2) Media performance (latest): spatial audio, low light, hdr; gaming: annual GGI + driver updates (3) 12GB+ RAM (4) Qualified SOC (flagship chip) (5) Quality in Field: Meet SLOs (e.g., crash rate) in 2026. Further enforced in 2027 (6) Quality at Launch: pass test suite on A17+ (7) OS: 5 upgrades, AVF, pKVM (8) Security: 6 yrs, quarterly"

Two items in that list barely appear in secondary coverage and both matter for planning. Requirement six is "pass test suite on A17+", an Android 17 baseline. Requirement seven demands five OS upgrades plus AVF and pKVM, which are virtualisation capabilities, not raw specification numbers. This is an OEM programme with quality and support commitments attached, not a spec sheet.

Requirement (Google's wording)	What it constrains	Practical read for an app team
"Nano models on device AI (integrate AI Core + Nano v3 or greater)"	Model generation on device	Your ML Kit GenAI code may run on a Nano v2 device but a Gemini Intelligence device is v3 or later
"12GB+ RAM"	Memory	The single clearest exclusion; mid-range 8GB flagships are out
"Qualified SOC (flagship chip)"	Silicon tier	Google qualifies specific SoCs; there is no published public list
"Quality at Launch: pass test suite on A17+"	OS version	An Android 17 baseline for the certification suite
"OS: 5 upgrades, AVF, pKVM"	Support and virtualisation	An OEM commitment, invisible to your code
"Security: 6 yrs, quarterly"	Patch cadence	An OEM commitment, invisible to your code
"Quality in Field: Meet SLOs (e.g., crash rate) in 2026. Further enforced in 2027"	Field reliability	Requirements tighten next year, so today's qualifying list is not stable

Worth flagging what Google did not say. Both the 12 May 2026 Gemini Intelligence announcement on the Google blog and the same-day Android Show developer post describe the features at length and contain no hardware requirements at all, beyond "starting with the latest Samsung Galaxy and Google Pixel phones this summer". If you have seen the 12GB figure attributed to a developer keynote, the attribution is wrong even though the number is right. Cite the android.com footnote.

Gemini Intelligence is not the same thing as Gemini Nano

This distinction decides your architecture, so be precise about it.

Gemini Intelligence is a consumer feature bundle on qualifying flagship hardware. Gemini Nano is the on-device model family your app talks to through the ML Kit GenAI APIs and the Prompt API. The populations are very different sizes. Google's Android Developers Blog stated on 21 July 2026 that "Gemini Nano was first introduced a few years ago, and is now running on over 140 million devices", and that "The latest version of the model, Gemini Nano 4, is built on the architecture foundation of the recently released Gemma 4 model."

Your app does not need a Gemini Intelligence device. It needs a device with AICore and a supported Nano model, which is a substantially wider set. The ML Kit GenAI device list covers Pixel 9 and Pixel 10 families, Samsung Galaxy S25, S26, Z Fold7 and Z TriFold, plus Honor, iQOO, Lenovo, Motorola, OnePlus, OPPO, POCO, realme, Sharp, vivo and Xiaomi models.

One detail contradicts a widely repeated claim: in Google's own Prompt API table, Pixel 9 is listed under nano-v3, while Galaxy Z Fold7 sits under nano-v2. Reporting that treats Pixel 9 as categorically excluded from the newer model does not match the developer documentation. Check the table before you write a device allowlist.

The API surface you actually code against

The ML Kit GenAI APIs expose six features, documented in Google's words as:

"Summarization: Summarize articles or chat conversations as a bulleted list. Proofreading: Polish short content by refining grammar and fixing spelling errors. Rewriting: Rewrite short messages in different tones or styles. Image description: Generate a short description of a given image. Speech recognition: Transcribe audio content to text. Prompt: Generate text content based on a custom text-only or multimodal prompt."

Maturity as documented: Summarization, Proofreading, Rewriting, Image description and Prompt are Beta; Speech recognition is Alpha. The dependencies are versioned accordingly, for example com.google.mlkit:genai-summarization:1.0.0-beta1 and com.google.mlkit:genai-prompt:1.0.0-beta3.

Underneath sits AICore. Google's Gemini Nano documentation describes it as "an Android system service that enables on-device execution of GenAI foundation models to facilitate features such as enhanced app functionality and improved user privacy by processing data locally", and adds an important isolation property: "AICore does not have direct internet access. All internet requests, including model downloads, are routed through the open-source Private Compute Services companion APK."

The constraints that will surprise you in review

These are all documented, and each one has ended a feature design we have seen proposed:

• "This API requires Android API level 26 or higher."
• "This API is not supported on devices with an unlocked bootloader."
• "GenAI API inference is permitted only when the app is the top foreground application", and violating it returns ErrorCode.BACKGROUND_USE_BLOCKED.
• "AICore enforces an inference quota per app", surfaced as ErrorCode.PER_APP_BATTERY_USE_QUOTA_EXCEEDED.
• Summarization input "must be under 4000 tokens (or approximately 3000 English words)", and ARTICLE input "must also be over 400 characters".
• Summarization supports "English, Japanese, and Korean" only.

The foreground restriction is the one that kills the most product ideas. Any design that says "summarise the user's inbox overnight" or "pre-generate descriptions in a WorkManager job" does not work on-device. That work goes to the cloud path or it does not happen.

The language list matters disproportionately for the Indian market, and we come back to it below.

Detect before you route

The correct pattern is not a device allowlist. It is a runtime capability check, because model availability changes as AICore downloads feature-specific adapters. Google's summarization guide documents four states:

// Check feature availability. Status will be one of the following:
// UNAVAILABLE, DOWNLOADABLE, DOWNLOADING, AVAILABLE
val featureStatus = summarizer.checkFeatureStatus().await()

if (featureStatus == FeatureStatus.DOWNLOADABLE) {
    summarizer.downloadFeature(object : DownloadCallback { /* ... */ })
}

Google notes that "If downloadFeature is not called, the first inference request will also trigger the feature to be downloaded if it's not already downloaded", and that "If Gemini Nano is already downloaded on the device, the feature-specific LoRA adapter model will be downloaded quickly." Calling downloadFeature explicitly is still better product design, because it lets you show progress instead of an unexplained delay on first use.

FeatureStatus	What it means	What the UI should do
AVAILABLE	Model and adapter are on device	Run inference; show the on-device path
DOWNLOADABLE	Supported device, adapter not fetched	Offer the feature, call downloadFeature, show progress
DOWNLOADING	Fetch in progress	Show progress; queue or route this request to cloud
UNAVAILABLE	Device cannot run it	Do not show an on-device promise; route to cloud or hide

getBaseModelName() returns the Gemini Nano version present on the device, which is how you tell a Nano v2 device from a Nano v3 device at runtime rather than by hardcoding a model list that goes stale every quarter.

The fallback Google actually recommends

Google's hybrid inference guidance is direct about why the cloud path exists: "Cloud models serve as a critical fallback when on-device models, such as Gemini Nano, are unavailable due to device hardware or OS constraints."

The product is the Firebase AI Logic Hybrid API, and the routing decision is a single enum.

InferenceMode	Behaviour (Google's documentation)	When to pick it
PREFER_ON_DEVICE	Attempts on-device, automatically falls back to cloud if unavailable or unsupported for the request	Default for privacy-sensitive text features with a wide device base
PREFER_IN_CLOUD	Cloud when online; on-device only when offline	Quality-sensitive features where the larger model matters
ONLY_ON_DEVICE	Throws if unavailable	Features that must never send data off device
ONLY_IN_CLOUD	Throws in all other cases	Capabilities Nano does not have at all

The Kotlin shape is small enough to read in one line of thought:

val model = Firebase
    .ai(backend = GenerativeBackend.Companion.googleAI())
    .generativeModel(
        modelName = "gemini-2.5-flash",
        onDeviceConfig = OnDeviceConfig(
            mode = InferenceMode.Companion.PREFER_ON_DEVICE
        )
    )

Google also sanctions custom routing rather than relying on the enum alone, on latency, battery and processor load, and query complexity, and cites GBoard's proofread and rewrite features and Kakao Mobility as production examples. In practice a custom router is the right call once you have more than one AI feature, because the correct mode differs per feature rather than per app.

What the fallback costs

The cloud path is not free, and this is where a feature that looked cheap in a demo becomes a line item. Google's Gemini API pricing publishes paid-tier rates per 1M tokens in USD. Figures below are as served on 3 August 2026.

Model and tier	Input per 1M tokens	Output per 1M tokens (including thinking tokens)
Gemini 3.1 Flash-Lite, standard	$0.25 text, image, video; $0.50 audio	$1.50
Gemini 3.1 Flash-Lite, Batch API	$0.125; $0.25 audio	$0.75
Gemini 3.1 Flash-Lite, priority	$0.45; $0.90 audio	$2.70
Gemini 3.5 Flash, standard	$1.50	$9.00
Gemini 3.5 Flash, Batch API	$0.75	$4.50

Google describes Gemini 3.1 Flash-Lite as "Our most cost-efficient model, optimized for high-volume agentic tasks, translation, and simple data processing", and the Batch API as a "50% cost reduction". Context caching for Flash-Lite is $0.025 and $0.05 per 1M tokens plus $1.00 per 1M tokens per hour of storage. Grounding with Google Search on Gemini 3 models gives 5,000 prompts per month free, then $14 per 1,000 grounded prompts; on Gemini 2.5 models it is 1,500 requests per day free, then $35 per 1,000 grounded prompts.

One deprecation to check in existing code: "Gemini 2.0 Flash-Lite is deprecated and has been shut down June 1, 2026."

The arithmetic that matters is per-user, not per-request. Applying the standard-tier rates in the table above, a summarisation feature averaging 2,000 input and 300 output tokens per invocation works out to roughly $0.00095 per call on Flash-Lite. At ten invocations per monthly active user, that is under a cent per user per month. At a hundred, it is roughly $0.095. The on-device path costs nothing per call, which is exactly why the routing decision is a financial one and not only a privacy one. The real cost is usually the tail of heavy users, not the median.

Sizing your own gap, because Google does not publish it

There is no official Android RAM distribution figure. The Android distribution dashboard footer states its snapshot covers a seven-day period ending 24 November 2025, and the version and RAM tables are served in an embedded frame rather than as page content. Google's own text points you elsewhere: RAM, SoC and Android version breakdowns are available through the Reach and devices report in Play Console, which is specific to your app and behind a login.

That is the right data source anyway. Fleet-wide averages tell you nothing useful; a fintech app in India and a design tool in the United States have different device mixes, and only your own report reflects yours.

The measurable figures Google does publish on that dashboard are graphics capabilities from the same November 2025 snapshot: Vulkan 1.1 at 62.09%, Vulkan 1.3 at 26.01%, Vulkan 1.4 at 0.67%, no Vulkan at 7.37%, and OpenGL ES 3.2 at 98.24% of handheld devices. Those are not substitutes for a RAM figure, and should not be quoted as one.

A tiering model that survives contact with the installed base

Three tiers, decided at runtime, is what we build.

Tier 1, on-device capable. checkFeatureStatus() returns AVAILABLE or DOWNLOADABLE. Run inference locally. Say so in the UI, because on-device processing is a real privacy claim and users respond to it. Handle BACKGROUND_USE_BLOCKED and the per-app quota error as normal states, not crashes.

Tier 2, cloud-routed. UNAVAILABLE, or the request exceeds Nano's limits, for example an article over 4,000 tokens or a language outside English, Japanese and Korean. Route to Flash-Lite. Budget for it per feature, and put a server-side rate limit in front of it before launch rather than after the first bill.

Tier 3, feature hidden. No network and no on-device model. Do not show a disabled button with a tooltip explaining what the user cannot have. Hide it. A feature that appears and fails reads worse than a feature that was never advertised.

The failure mode to design against is a device that moves between tiers. A user on a supported handset who has never triggered the download sits in DOWNLOADABLE, not AVAILABLE, and will move to AVAILABLE later. Your UI has to tolerate that transition without a restart. This is the same class of capability-tiering problem as building for foldable and large-screen adaptive layouts, where a static device check produces an app that is wrong half the time.

India-specific considerations

Three points change the calculation for an India-facing Android product.

The language constraint is the binding one. ML Kit summarization supports English, Japanese and Korean. No Indic language is on that list. For an app serving Hindi, Tamil, Bengali, Marathi or Telugu content, the on-device path does not exist today for summarisation regardless of how much RAM the handset has, and the cloud route is the only route. Budget for it as the primary path, not the fallback.

The 12GB RAM floor lands harder on the Indian device mix than on a United States or Western European mix, where flagship share is higher. Since Google publishes no distribution data, treat that as a hypothesis to test against your own Play Console Reach and devices report before you commit a roadmap to on-device inference.

Under the Digital Personal Data Protection Act 2023, the routing decision carries a data-handling consequence that a purely technical framing misses. On-device inference through AICore keeps content local, and Google documents that AICore has no direct internet access. A PREFER_ON_DEVICE mode that silently falls back to a cloud model sends the same content to a Google endpoint. If your privacy notice describes AI processing as on-device, PREFER_ON_DEVICE is not the mode that claim describes; ONLY_ON_DEVICE is. Decide which promise you are making before you pick the enum.

For teams already planning the Android target API 36 Play Store deadline, the capability-detection work here fits into the same release train rather than competing with it.

What is confirmed about Pixel 11, and what is not

Google's own store page confirms the date and nothing else. The page states "The next generation. Google Pixel 11. August 12.", "Tune in live to see our newest devices drop at the Made by Google event.", "Google Pixel 11. Pre-order August 12." and "Tune in live from NYC."

Google has published no chip, no RAM figure, no display, no camera, no price and no model list, and the page does not name a Pro, Pro XL or Fold variant. Every specification circulating for Pixel 11, including chip generation, process node, storage tier, modem and colourways, is leak reporting rather than a Google statement.

There is a real open question worth watching rather than assuming. The Gemini Intelligence footnote requires passing a test suite on Android 17 or later, and Pixel 11 launches on 12 August 2026, but Google has not stated that Pixel 11 ships with Android 17 or that it qualifies for Gemini Intelligence. Do not put that assumption in a roadmap document.

One narrower device requirement Google has published, for a single feature: screen automation is documented as needing "Pixel 10, Pixel 10 Pro, Pixel 10 Pro XL, or Samsung Galaxy S26 Series", with the user aged 18 or over and located in the US or Korea, and Pixel 10 devices not supported in Korea. That is the shape of these requirements in practice, feature by feature and region by region, which is another argument for runtime detection over a hardcoded matrix. The same reasoning applies to agent-ready Android apps built on Gemini and App Actions, where the capability surface differs across devices in the same product line.

FAQ

Does Gemini Intelligence really require 12GB of RAM?

Yes. Google's own requirement list on android.com names "12GB+ RAM" alongside a "Qualified SOC (flagship chip)" and "Nano v3 or greater". The same footnote adds an Android 17 test suite requirement, five OS upgrades with AVF and pKVM, and six years of quarterly security updates.

Do my app's AI features need a Gemini Intelligence device?

No. Gemini Intelligence is a consumer feature bundle on qualifying flagship hardware. Your app talks to Gemini Nano through the ML Kit GenAI APIs, which run on a much wider device set. Google stated on 21 July 2026 that Gemini Nano runs on over 140 million devices.

Which ML Kit GenAI features are available on Android?

Google documents six: summarization, proofreading, rewriting, image description, speech recognition and a general prompt API. Summarization, proofreading, rewriting, image description and prompt are labelled Beta, while speech recognition is Alpha. All require Android API level 26 or higher and are unsupported on unlocked bootloaders.

How do I check at runtime whether on-device AI will work?

Call checkFeatureStatus(), which returns one of four documented values: UNAVAILABLE, DOWNLOADABLE, DOWNLOADING or AVAILABLE. On DOWNLOADABLE, call downloadFeature to fetch the adapter with visible progress rather than blocking on first inference. Use getBaseModelName() to read the Gemini Nano version present, rather than hardcoding a device allowlist that goes stale.

What does Google recommend when a device cannot run on-device inference?

The Firebase AI Logic Hybrid API. Google states that cloud models serve as a critical fallback when on-device models are unavailable due to device hardware or OS constraints. The PREFER_ON_DEVICE inference mode attempts on-device first and falls back to a cloud model automatically when it is unsupported.

How much does the cloud fallback cost?

Gemini 3.1 Flash-Lite is priced at $0.25 per 1M input tokens for text, image and video, and $1.50 per 1M output tokens on the paid standard tier. The Batch API halves that to $0.125 and $0.75. Gemini 3.5 Flash standard runs $1.50 input and $9.00 output.

Can on-device AI run in the background on Android?

No. Google documents that GenAI API inference is permitted only when the app is the top foreground application, and background attempts return ErrorCode.BACKGROUND_USE_BLOCKED. AICore also enforces a per-app inference quota. Any batch or overnight processing design has to use the cloud path instead.

Do the on-device APIs support Indian languages?

Not for summarization. Google documents support for English, Japanese and Korean only. For Hindi, Tamil, Bengali or other Indic content the on-device summarization path is unavailable regardless of device RAM, so cloud inference becomes the primary route rather than a fallback for those users.

How eCorpIT can help

eCorpIT is a Gurugram-based technology consulting organisation whose senior engineering teams build Android products where AI capability varies across the installed base. We run the Play Console Reach and devices analysis to size your actual gap, design the runtime detection and routing layer rather than a device allowlist, and model per-user cloud inference cost before the feature ships. We are CMMI Level 5, MSME certified and ISO 27001:2022 certified, and we design applications aligned with DPDP requirements where inference routing affects where user content is processed. To scope an on-device AI feature against your own device mix, contact us.

References

1. Gemini Intelligence, Google, android.com. Footnote 1 carries the full device requirement list.
2. ML Kit GenAI APIs, Google for Developers, last updated 21 July 2026.
3. Summarization with ML Kit GenAI on Android, Google for Developers, last updated 19 June 2026.
4. Gemini Nano and AICore, Android Developers, last updated 2 April 2026.
5. Hybrid inference on Android, Android Developers, last updated 1 June 2026.
6. Android on-device inference update, Android Developers Blog, 21 July 2026.
7. The Android Show: Developers' Cut 2026, Android Developers Blog, 12 May 2026.
8. Gemini Intelligence announcement, The Keyword, Google, 12 May 2026.
9. Gemini API pricing, Google AI for Developers. Figures as served 3 August 2026.
10. Get started with the Firebase AI Logic hybrid API on Android, Firebase documentation.
11. Android distribution dashboard, Android Developers. Snapshot period ending 24 November 2025.
12. Use screen automation on your Pixel phone, Google Pixel Help.
13. Google Pixel 11, Google Store. Confirms the 12 August 2026 Made by Google date.

Last updated: 3 August 2026.