On-device AI is now a buying criterion for phones, laptops and browsers, but each platform sets its own hardware bar, splits work differently between the device and the cloud, and makes different privacy promises. This hub sets out, field by field, what five on-device AI platforms require and offer: Apple Intelligence, Chrome built-in AI, Copilot+ PCs, Galaxy AI and Gemini Nano on Android. It covers supported hardware, storage and memory, features, the device and cloud split, privacy architecture, developer APIs, enterprise controls and regional limits, all from the providers' own pages. It is for buyers, IT teams, developers and anyone deciding which devices can run which AI features. Platforms are listed alphabetically; order is never a ranking. 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
- Every platform sets a different hardware bar. Copilot+ PCs need an NPU of more than 40 TOPS, Apple Intelligence needs an iPhone 15 Pro or later or an M1 or later iPad or Mac, and Chrome built-in AI needs 22 GB of free storage plus a GPU with more than 4 GB of VRAM or 16 GB of RAM.
- "On-device" rarely means everything stays on the device. Apple sends some requests to Private Cloud Compute, Samsung states Photo Assist and Creative Studio need a network connection, and Android's Firebase AI Logic hybrid API can fall back to Gemini in the cloud.
- Privacy architectures are documented, and they differ. Apple states no user data is retained in Private Cloud Compute after a response; Android's AICore stores no record of inputs or outputs; Windows Recall keeps encrypted snapshots locally behind Windows Hello.
- Free today, with paid tiers signalled. Apple says expanded access to server-side features will be available for a fee in future, and Samsung says future advanced Galaxy AI features may be offered through a paid subscription.
- Regional and language limits matter. Apple Intelligence does not currently work on devices bought in mainland China, Chrome's models support five languages from Chrome 149, and Copilot+ experiences vary by device and market.
What on-device AI means
On-device AI runs a model on the phone, PC or browser's own processor (CPU, GPU or a dedicated neural processing unit, NPU) instead of sending the request to a data centre. The practical benefits providers cite are offline use, lower latency, lower cost and keeping personal data local. The trade-off is model size: on-device models are smaller than cloud models, so platforms either limit what runs locally or hand larger tasks to a cloud service. Google's Android documentation names token limits, hardware constraints and capability boundaries as Gemini Nano's limits, and recommends hybrid inference for those cases.
An AI PC, in Microsoft's definition, is a Copilot+ PC: a Windows 11 device with an NPU that can perform more than 40 trillion operations per second (TOPS). TOPS figures published by chip makers are measured differently and are not directly comparable across vendors; see What Is an NPU? TOPS, AI PC Chips and What the Numbers Mean.
Platforms at a glance
As of 2026-11-15. Platforms alphabetical. Identical fields for each.
| Field | Apple Intelligence | Chrome built-in AI | Copilot+ PCs | Galaxy AI | Gemini Nano (Android) |
|---|---|---|---|---|---|
| Owner | Apple | Microsoft (with PC makers) | Samsung | ||
| Type | Personal AI system across Apple devices | Browser-managed models and web APIs | Windows 11 PC class with NPU | Phone AI feature suite | On-device foundation model in Android |
| Model named by provider | AFM 3 Cloud models (in Shortcuts, server side) | Gemini Nano | Not stated on the Copilot+ developer page | Not stated in the S26 announcement | Gemini Nano |
| Runs where | On device, plus Private Cloud Compute for server features | In the browser on the device | On the device NPU | Mix; some features need a network connection | In the AICore system service on the device |
| Price | Included; server features have daily limits, expanded access for a fee planned | Not stated by Google as a charge | Included with the device; some Click to Do actions need a subscription | Free; future advanced features may be paid | Not stated by Google as a charge; ML Kit GenAI API terms apply |
| Main source | Apple Support | Chrome for Developers | Microsoft Learn | Samsung Newsroom | Android Developers |
Hardware and device requirements
As of 2026-11-15. Platforms alphabetical.
| Platform | Supported devices or chips | Memory and storage | Operating system |
|---|---|---|---|
| Apple Intelligence | iPhone Duo, iPhone 16 and later, iPhone 15 Pro and Pro Max, iPhone Air; iPad mini (A17 Pro and A20 Pro) and iPads with M1 or later; MacBook Neo (A18 Pro) and Macs with M1 or later; Apple Vision Pro; Apple Watch Series 9 or later, Ultra 2 or later and SE 3 when paired with a supported iPhone | Up to 14 GB storage on iPhone Duo, iPhone 17 Pro and later, iPhone Air, M4 or later iPads and M3 or later Macs with at least 12 GB unified memory, and Vision Pro (M5); up to 8 GB on other supported models; up to 1.5 GB on Apple Watch | iOS 27, iPadOS 27, macOS 27 Golden Gate, visionOS 27, watchOS 27 |
| Chrome built-in AI | Desktop computers and Chromebook Plus devices | At least 22 GB free on the Chrome profile volume; GPU with more than 4 GB VRAM, or CPU with 16 GB RAM and 4 or more cores; audio input to the Prompt API needs a GPU | Windows 10 or 11, macOS 13 or later, Linux, ChromeOS on Chromebook Plus (platform 16389.0.0 onwards); not Android or iOS |
| Copilot+ PCs | NPU of 40+ TOPS; Microsoft lists Snapdragon X Elite, AMD Ryzen AI 300 series and Intel Core Ultra 200V series | Recall additionally needs 16 GB RAM, 8 logical processors, 256 GB storage with 50 GB free | Windows 11 |
| Galaxy AI | Galaxy S26 series, unveiled 25 February 2026 | Not stated in the S26 announcement | Not stated in the S26 announcement |
| Gemini Nano (Android) | Android devices that provide AICore | Not stated on the Gemini Nano overview page | Android |
Device detail sits on each platform's own explainer: Apple Intelligence Explained, Copilot+ PCs Explained, Galaxy AI Explained and Gemini Nano Explained.
Features and the device to cloud split
Apple Intelligence
Apple lists Siri AI, Intelligent Photo Editing Tools, Image Playground and AFM 3 Cloud models in Shortcuts as server-side features with daily usage limits that vary by feature, request complexity and demand. Some Apple Watch features need an iPhone nearby. In the EU, Apple states that advanced photo editing, the reimagined Image Playground and intelligent browsing are available on the listed OS 27 devices.
Chrome built-in AI
Chrome provides seven APIs to websites and extensions: Prompt (general purpose), Summarizer, Writer, Rewriter, Proofreader, Translator and Language Detector. Google states some are in Chrome stable, some in origin trials and some limited to its Early Preview Program. The Translator and Language Detector APIs work on desktop Chrome only.
Copilot+ PCs
Microsoft names Click to Do and improved Windows search as Copilot+ experiences, and Recall as a Copilot+ PC feature. It notes that Click to Do actions other than image actions vary by device, region, language and character set, and that improved search works with specific file formats and select languages.
Galaxy AI
Samsung's S26 announcement lists Now Nudge (contextual suggestions, Samsung account, certain apps), Circle to Search with Google, Photo Assist (object removal and scene changes, network and Samsung account required, visible watermark on saved images), Creative Studio (sticker sets from prompts, network and account required) and Bixby suggestions. Gemini (Google account, network, select languages) and Perplexity (pre-installed, network and Perplexity account) are part of the assistant line-up.
Gemini Nano (Android)
Google's ML Kit GenAI APIs expose Gemini Nano for direct prompts (GenerativeModel), agent loops with tool calling (AgentRunner), speech recognition, speech translation, text to speech, image description, summarisation, proofreading, rewriting and semantic search with embeddings. Prompts run locally with no server calls; apps that need more can use the Firebase AI Logic hybrid API to fall back to Gemini Flash or Gemini Pro in the cloud.
Privacy architecture
As of 2026-11-15. Only what each provider documents. Platforms alphabetical.
| Platform | Local processing | Cloud processing | Stated retention |
|---|---|---|---|
| Apple Intelligence | On-device models for local features | Private Cloud Compute, designed for stateless computation, enforceable guarantees, no privileged runtime access, non-targetability and verifiable transparency | Apple states no user data is retained in any form after the response is returned; node data keys are randomised on every reboot |
| Chrome built-in AI | Gemini Nano in the browser after a one-time download | Not used for built-in AI calls; server-side generative AI is a separate, documented option | Google states no data is sent to Google or third parties when using the model |
| Copilot+ PCs (Recall) | Snapshots stored and analysed locally | Not used for Recall; snapshots are not sent to Microsoft | Snapshots always encrypted; keys protected by the TPM, tied to Windows Hello Enhanced Sign-in Security and used in a VBS enclave; retention set by policy (30 to 180 days on Enterprise and Education) |
| Galaxy AI | Not broken down in the S26 announcement | Photo Assist, Creative Studio, Gemini and Perplexity require a network connection | Not stated in the S26 announcement |
| Gemini Nano (Android) | AICore runs requests locally; no direct internet access, downloads go through Private Compute Services | Optional app-controlled fallback through Firebase AI Logic, protected with Firebase App Check | AICore isolates each request and keeps no record of inputs or outputs after processing |
For the security assessment of these architectures, threat models and data leakage risks, see the On-Device AI Security Guide. This hub does not repeat it.
Developer APIs
| Platform | APIs and frameworks named by the provider | Fallback behaviour |
|---|---|---|
| Apple Intelligence | AFM 3 Cloud models available in Shortcuts (per Apple Support); developer framework detail is on the Apple Intelligence explainer | Server features subject to daily limits |
| Chrome built-in AI | Prompt, Summarizer, Writer, Rewriter, Proofreader, Translator, Language Detector APIs | Model removed if free space falls below 10 GB and re-downloaded when requirements are met |
| Copilot+ PCs | Microsoft Foundry on Windows APIs (in the Windows App SDK); Windows ML on ONNX Runtime with execution providers such as QNN and OpenVINO; Olive for model optimisation | Windows ML falls back to GPU or CPU |
| Galaxy AI | Not stated in the S26 announcement | Not stated in the S26 announcement |
| Gemini Nano (Android) | ML Kit GenAI APIs on AICore (GenerativeModel, AgentRunner, speech, summarisation, proofreading, rewriting, image description, semantic search) | Firebase AI Logic hybrid API to cloud Gemini |
Enterprise management and identity
- Copilot+ PCs: Microsoft states Recall is disabled and removed by default on managed devices, and admins cannot start saving snapshots for users. Policies include Allow Recall to be enabled (AllowRecallEnablement), Turn off saving snapshots for Recall (DisableAIDataAnalysis), maximum storage and duration, and app and URI deny lists (SetDenyAppListForRecall, SetDenyUriListForRecall).
- Apple Intelligence: management controls for Apple devices are covered in Apple Intelligence Enterprise Security.
- Galaxy AI: Samsung states several features require a Samsung account, Gemini requires a Google Account and Perplexity a Perplexity account, so account policy is part of deployment.
- Gemini Nano and Chrome: developers own client safety; Google states the developer is responsible for the safety of the API client and app experience, and recommends App Check for any cloud fallback.
Identity is the common thread: Recall is gated by Windows Hello biometrics, Galaxy and Gemini features are gated by consumer accounts, and cloud fallbacks rely on app attestation. Treat these accounts and app credentials as part of the device's identity estate; see the Mobile Device Security Guide.
Languages and regions
| Platform | Languages stated | Regional limits stated |
|---|---|---|
| Apple Intelligence | iOS, iPadOS, macOS: English, Danish, Dutch, French, German, Italian, Norwegian, Portuguese, Spanish, Swedish, Turkish, Vietnamese, Chinese (Simplified and Traditional), Japanese, Korean; device and Siri language must match | Not currently working on devices bought in mainland China or with a mainland China account there |
| Chrome built-in AI | English, Spanish, Japanese, German, French from Chrome 149 | Not stated on the get-started page |
| Copilot+ PCs | Improved search optimised for select languages | Experiences vary by device and market |
| Galaxy AI | Gemini available in select languages | Not stated in the S26 announcement |
| Gemini Nano (Android) | Not stated on the overview page | Not stated on the overview page |
How to choose: factual criteria
- Start from the devices you already run. Each platform is tied to its own hardware and OS, so the fleet usually decides the platform.
- Check the hardware bar for the features you want. A Windows laptop without a 40+ TOPS NPU is not a Copilot+ PC, and Recall adds RAM and storage minimums on top.
- Map which features leave the device. Use the privacy table to see which features use cloud processing and what the provider states about retention.
- Check languages and regions for every market where staff or customers use the device.
- Plan for paid tiers. Apple and Samsung have both signalled future paid access for some features.
- Confirm admin controls before enabling features on managed devices. A side by side of the three phone platforms is in Apple Intelligence vs Galaxy AI vs Gemini on Pixel.
Practitioner checklist
- Inventory devices against each platform's supported hardware list and storage needs.
- Record which AI features on each device use cloud processing.
- Set Recall policies (AllowRecallEnablement, DisableAIDataAnalysis, deny lists) before Copilot+ PCs reach users.
- Decide which consumer accounts (Samsung, Google, Perplexity) are allowed on managed phones.
- Check language and regional availability for each market you operate in.
- For apps using Gemini Nano or Chrome APIs, design a fallback and protect any cloud path with app attestation.
- Re-check provider pages before purchase decisions: device lists and paid tiers change.
Standards and references
- Apple Support: How to get Apple Intelligence
- Apple Security Research: Private Cloud Compute
- Chrome for Developers: Get started with built-in AI
- Chrome for Developers: Built-in AI
- Microsoft Learn: Copilot+ PCs developer guide
- Microsoft Learn: Manage Recall
- Samsung Newsroom: Galaxy Unpacked 2026, Galaxy S26 series
- Android Developers: Gemini Nano
Related NHI Mgmt Group resources: On-Device AI Security Guide · What Is an NPU? TOPS, AI PC Chips and What the Numbers Mean · Apple Intelligence vs Galaxy AI vs Gemini on Pixel · Apple Intelligence Enterprise Security