Apple Intelligence Local LLM: On-Device Models vs Private Cloud Compute

Apple Intelligence runs partly on your device and partly in Apple’s Private Cloud Compute (PCC): some requests stay on the device, more demanding ones can move to PCC, and selected features can send requests to ChatGPT under user control (normal request confirmations can be switched off; files and photos still require separate permission). The on-device language model Apple documented in 2025 has roughly 3 billion parameters; the 2026 generation adds AFM 3 Core Advanced with 20 billion total parameters, of which 1–4 billion are active per request. Apple now lists up to 8GB or 14GB of device storage, depending on model class; the exact automatic routing rules between the device and PCC remain unpublished (checked October 8, 2026).

Sources rechecked October 13, 2026. The next generation of Apple Intelligence was released September 14, 2026 with iOS 27, iPadOS 27 and macOS 27. Siri AI began rolling out as an English-language beta, rather than making every feature, language and device capability generally available at once. Apple now also applies daily usage limits to some server-backed features; those limits depend on the feature and usage. The 2025/June-2026 architectural analysis below remains useful, but its preview-era status has been superseded. Apple release · PCC expansion

Apple Intelligence is frequently described as “Apple’s local LLM,” but that label hides most of the architecture. It is not one model and it is not simply a renamed Siri. Apple combines several on-device language, vision, audio, and classification models with operating-system services, app tools, Private Cloud Compute (PCC), and optional third-party integrations such as ChatGPT. ¹²³

How to read these claims

• Officially confirmed: documented by Apple, a platform vendor, or the relevant open-source project.
• Released or beta: the September 14 OS rollout is public; Siri AI and individual feature/language paths may still be beta or limited.
• Architecture-level conclusion: a reasonable inference, not a disclosed internal Apple implementation detail.
• Not published: Apple has not provided the necessary data.

What matters in practice

• Apple Intelligence is a hybrid AI architecture. Some requests stay on the device, more demanding work can move to PCC, and selected experiences can invoke an external provider. ¹²³
• Apple’s 2025 technical report described a highly compressed on-device language model with roughly 3 billion parameters. In June 2026, Apple introduced AFM 3 Core and the sparse multimodal AFM 3 Core Advanced, which has 20 billion total parameters but activates only 1–4 billion for a request. ⁴⁵
• AFM 3 Core Advanced keeps the full expert set in NAND flash and loads a selected subset into active memory. Apple limits this path to its most capable silicon configurations. ⁴
• Apple does not publish the complete automatic routing policy between the device, PCC, and external models. A feature name alone therefore cannot prove where every request was processed. ²⁶
• For the macOS/iOS 27 generation, Apple lists up to 8GB of storage for several previously supported devices and up to 14GB on M3-or-newer Macs with at least 12GB unified memory. These are storage amounts, not measured local-model RAM usage. ⁷
• The Foundation Models framework exposes Apple’s system model through native Swift APIs. The 2026 generation expands the framework with multimodal input, model-provider protocols, PCC access for eligible developers, evaluation tools, Dynamic Profiles, a command-line tool, and a Python SDK. ⁸⁹¹⁰
• MLX, Core ML, Core AI, and llama.cpp can all run local AI workloads on Apple hardware, but they are not interchangeable with Apple Intelligence itself. ^11121314

What is Apple Intelligence?

Apple calls Apple Intelligence a personal intelligence system built into its operating systems. In practice, it is a coordinated platform rather than a single general-purpose chatbot. The platform connects generative models to app actions and relevant personal context while applying data-minimization, permission, safety, and routing rules. ¹¹⁵

A useful architectural breakdown is:

1. User-facing experiences: Writing Tools, message and notification summaries, Siri, Genmoji, Image Playground, visual intelligence, smart actions in Shortcuts, translation, and app-specific features.
2. System orchestration: operating-system services assemble the required context, expose approved tools, enforce permissions, and choose an execution path.
3. On-device models: language, vision, speech, ranking, and other specialized models that can execute locally.
4. Private Cloud Compute: larger Apple models for requests that exceed the local path’s capability or resource envelope.
5. External model integrations: a separate, optional path such as ChatGPT, used only in supported experiences and subject to user controls. ²³¹⁶

This distinction matters. It is accurate to say that Apple Intelligence has extensive on-device processing. It is inaccurate to claim that the entire system always runs locally.

The diagram separates on-device models, system orchestration, Private Cloud Compute, and an optional external service such as ChatGPT. Apple's complete per-request routing policy is not public; the PCC documentation describes the server-side path.

What does “Apple Intelligence local LLM” mean?

The search phrase can refer to three very different systems.

Apple’s operating-system foundation model

Apple ships and manages an on-device model as part of the OS. Apple Intelligence and the Foundation Models framework can use it without an app bundling its own multi-gigabyte weights. Developers receive an API, not an unrestricted copy of the production model weights. Apple controls updates, compression, safety behavior, language availability, and hardware compatibility. ⁵⁸

Publicly documented Apple Foundation Models

Apple publishes research articles and technical reports describing model families, optimization techniques, evaluations, and safety work. Those reports improve transparency, but they do not turn the production Apple Intelligence models into downloadable open-weight releases. ⁴⁵¹⁷

User-installed open models on a Mac

A Mac owner can separately run Llama, Qwen, Mistral, Gemma, or other open models through MLX, llama.cpp, Core ML, or third-party applications. These deployments can offer user-selected context windows, sampling parameters, retrieval pipelines, adapters, and private local endpoints. They do not automatically become Siri’s model or gain Apple Intelligence’s privileged operating-system context. ¹³¹⁴

Apple’s on-device model architecture

The 2025 on-device model

Apple’s 2025 foundation-model report describes a multilingual and multimodal on-device language model with approximately 3 billion parameters. It was optimized for Apple silicon using quantization-aware training. Apple reported roughly 2-bit main weights, a 4-bit embedding table, and an 8-bit KV cache in the documented configuration. ⁵

The model also used KV-cache sharing between transformer blocks. Apple reported a 37.5% reduction in KV-cache memory and a similar reduction in time to first token for the evaluated architecture. These are model-design results from Apple’s report, not a universal benchmark for every iPhone, iPad, or Mac. ⁵

For visual understanding, the 2025 on-device family used an approximately 300-million-parameter ViTDet-L-based encoder, while the server family used a larger visual encoder. ⁵

Apple positioned the small system model for focused tasks such as summarization, extraction, text understanding, rewriting, short conversations, and tool-assisted app workflows. The report explicitly does not support treating it as a full replacement for a large, knowledge-heavy cloud chatbot. ⁵

Apple Foundation Models 3 in 2026

On June 8, 2026, Apple announced a five-model family created in collaboration with Google: ⁴

Model	Execution location	Apple’s stated role	Status after September 14, 2026 release
AFM 3 Core	On-device	next-generation dense ~3B model	Introduced with the 27-generation platforms
AFM 3 Core Advanced	On-device	multimodal sparse model; 20B total, 1–4B active	Device and feature dependent; not supported by every Apple silicon Mac
AFM 3 Cloud	Private Cloud Compute	server model optimized for speed and efficiency	Used by selected features and Shortcuts; daily quotas may apply
ADM 3 Cloud (Image)	Private Cloud Compute	image generation and editing	Powers new cloud image paths; availability varies by feature
AFM 3 Cloud Pro	Extended PCC	reasoning and agentic tool use	Advanced PCC model; production routing shares remain undisclosed

AFM 3 Core Advanced addresses a central constraint of consumer-device inference: keeping a large model’s full weights in DRAM. Apple says the full expert set is stored in NAND flash. A lightweight dense routing block chooses a fixed group of experts during initial prompt processing and periodically reselects experts during generation. Shared experts remain active while routed experts are loaded only when needed. ⁴

This is not conventional flash swapping on every generated token. Apple designed the router specifically to avoid the bandwidth cost of standard token-level mixture-of-experts loading. The result is a model with 20 billion total parameters but a much smaller active parameter set for any one request.

What Apple has not disclosed

Apple has not published a durable production table containing:

• one universal context window covering all internal on-device models and consumer features,
• per-model download sizes,
• peak and steady-state RAM use by feature,
• token-per-second results for each supported chip,
• a complete operator map across CPU, GPU, and Neural Engine,
• the full automatic routing policy between local, PCC, and external models.

For the specifically documented Foundation Models developer APIs, Apple lists a 4K context window for the on-device system model and 32K for the PCC model. These do not establish universal limits across all internal Apple Intelligence features. The PCC option requires a network connection and is subject to per-user daily quotas. WWDC26: PCC developer model.

Supported devices and hardware requirements

Current system availability and device tiers

As of the September 14, 2026 release, Apple lists Apple Intelligence for iPhone 15 Pro/Pro Max and iPhone 16 or newer; iPad mini (A17 Pro), iPads with M1 or newer; MacBook Neo (A18 Pro) and M1-or-newer Macs; and Apple Vision Pro. On Apple Watch Series 9+, Ultra 2+, or SE 3, eligible features require a compatible iPhone nearby. Older Series 6/SE 2 watches are not on Apple’s current Apple Intelligence support list. Specific features, including the most advanced on-device model, have stricter chip/RAM requirements.

Baseline support is not a promise of every new feature. Apple lists eligible device and Siri languages, but individual functions have region and language restrictions. Cloud-intensive features including Siri AI, Image Playground and AFM 3 Cloud in Shortcuts can have daily request limits. Check the current September 2026 availability notes and Apple Intelligence support page before buying a device.

United States: Siri AI beta availability is not the EU iPhone restriction

Sources rechecked October 13, 2026: Apple’s current U.S. documentation covers iOS 27, iPadOS 27, macOS 27, watchOS 27 and visionOS 27. Siri AI beta launched in English with the September software release on eligible devices. Feature support depends on the exact hardware, operating system, language and region. An EU-specific delay for Siri AI on iPhone and iPad should not be presented as a limitation for U.S. devices. Apple U.S. support.

The fact that a U.S. feature is available is not a promise of U.S.-only data residency or local inference. Shortcuts and optional integrations can choose a cloud model; check the actual selected data path and account rules before sharing sensitive information.

Higher requirements for selected 2026 capabilities

Apple attaches stricter hardware notes to several 27-generation features. For example, selected new Siri voice and higher-accuracy dictation capabilities require devices such as iPhone 17 Pro/Pro Max or iPhone Air, M4-class iPads with at least 12 GB unified memory, or M3-class Macs with at least 12 GB unified memory. These are feature-specific requirements, not a replacement for the general Apple Intelligence compatibility list. ¹⁸

Support does not mean that every model runs locally on every device. Apple is creating tiers inside the supported-device family: a device may support Apple Intelligence and the compact AFM 3 Core path while lacking the memory, bandwidth, or thermal budget for AFM 3 Core Advanced. There is no single published formula that maps chip specifications to model access; eligibility is documented at the feature level.

The diagram separates baseline Apple Intelligence compatibility from higher hardware requirements for selected 2026 features; Apple lists those thresholds in its support documentation.

Which Apple Intelligence features run on-device?

There is no complete public feature-to-model matrix. The same interface can use a different route depending on request complexity, device generation, language, software version, and user choice. The conservative table below avoids turning marketing categories into unsupported execution claims.

Feature or execution path	On-device	PCC	External service	Evidence-based assessment
Foundation Models framework with the system on-device model	yes	no for this explicitly selected path	no	Apple describes it as embedded in the OS and available offline. 8
Summarization, extraction, and rewriting through that local framework path	yes	only if the app chooses another provider	possible through the provider protocol	The app’s selected model is the deciding factor. 8910
Writing Tools in system apps	often	possible for harder tasks	ChatGPT can be used for Compose or an explicit integration	Apple does not guarantee that every Writing Tools request uses one fixed local model. 73
Mail, Messages, and notification summaries	possible	possible	not normally needed	The features are documented; the per-request execution map is not. 71
Siri	partly	yes for complex requests	ChatGPT with consent/configuration	Siri is an orchestrator and interface, not one model. 183
Live Translation	Apple highlights on-device execution for key paths	future implementations may vary	not normally	Offline and privacy benefits are official; the full model chain is feature- and device-dependent. 19
Genmoji	model- and version-dependent	the 2026 ADM 3 Cloud model can power Genmoji	not normally	It should not be labeled universally local. 4
Image Playground	earlier/lightweight styles can use local components	2026 high-quality and photorealistic generation uses PCC	separate ChatGPT image generation may be offered	The selected style and OS version matter. 49
2026 enhanced dictation	local on listed high-end devices	not for Apple’s stated on-device path	no	AFM 3 Core Advanced is explicitly an on-device model. 184
Visual intelligence	local preprocessing can occur	Apple server paths can occur	ChatGPT and Google Search can be deliberately invoked	The action selected by the user determines the data flow. 7
Shortcuts “Use Model” action	selectable	selectable	ChatGPT selectable	This is one of Apple’s clearest explicit model-routing controls. 20

Why a permanent local-versus-cloud checklist would be misleading

Apple Intelligence is task-oriented. “Summarize” may refer to a short email, a long set of notes, an image-plus-text request, or one step in an agentic workflow. The visible feature name therefore does not uniquely identify the model.

A reproducible test should record:

• device and chip,
• OS build,
• language and region,
• app and feature,
• whether networking was available,
• any visible model selection,
• input length and modality.

Without those details, a statement such as “Genmoji is local” or “Siri uses the cloud” is too broad to remain accurate across platform generations.

The flow shows minimization before a request enters an on-device or PCC route; only task-relevant context should enter the selected route. Apple's PCC architecture describes the server-side path.

How does Apple choose between local processing and PCC?

Apple’s public description is high-level: process on-device when possible; use a larger PCC model when the request requires more capability. ²¹⁵

Apple has not published the full dispatcher. The following factors are technically plausible but should not be presented as the complete internal algorithm:

• capability and modality required,
• context size and task complexity,
• which model tier the device supports,
• available memory, power, and thermal state,
• language support,
• safety or policy requirements,
• explicit user or developer selection.

For PCC, the device can verify that the destination is running an approved, publicly attested software image. The request is encrypted to an eligible PCC node. Apple also describes relay and privacy mechanisms intended to prevent the inference service from directly linking an IP address to request content. ²

The decision tree shows the possible local, PCC, and external routes; Apple does not publish the exact internal thresholds.

Private Cloud Compute architecture

PCC is Apple’s attempt to extend some of the device’s privacy properties to server-scale inference. Apple defined five core requirements: ²

1. Stateless user-data processing: request data must not be retained after the response.
2. Enforceable guarantees: privacy must be implemented in hardware and software rather than relying only on policy.
3. No privileged runtime access: administrators must not have a special data-access path into production inference.
4. Non-targetability: attackers should not be able to route one user to a specially modified node.
5. Verifiable transparency: researchers should be able to inspect the software associated with production deployments.

Original PCC nodes

Apple’s 2024 design used Apple silicon, Secure Enclave, Secure Boot, a hardened operating-system subset, strict code signing, and a minimal runtime surface. Apple says the production environment avoids general-purpose remote shells and conventional logging paths that could expose user request data. Public transparency records and published software artifacts allow devices and researchers to verify deployment claims. ²

PCC expansion to Google Cloud and NVIDIA GPUs

In June 2026, Apple announced that the most demanding Apple Intelligence workloads could run on NVIDIA GPUs in Google Cloud under an expanded PCC model. Apple cites NVIDIA Confidential Computing, Intel TDX, Google Titan, independent trust roots, attested keys, an append-only ledger, and published binaries as elements of the new chain of trust. The capacity was being introduced gradually during the summer preview. ⁶

This changes an important architectural assumption. PCC no longer means only Apple-silicon servers in Apple-owned data centers. Apple’s position is that the same privacy requirements can be maintained through a multi-party confidential-computing design. That claim is official; the architecture should still be evaluated through ongoing independent security research as deployment matures.

The security diagram shows encrypted requests, attested software, stateless processing, and public transparency artifacts described by Apple Security Research.

Privacy models: on-device, PCC, and ChatGPT are not equivalent

Route	Processing location	Model operator	Retention claim	User control
On-device system model	iPhone, iPad, or Mac	Apple’s OS model	inference input need not leave the device	sometimes implicit; explicit in developer APIs
Private Cloud Compute	attested PCC environment; in 2026 potentially extended Google Cloud infrastructure	Apple-controlled PCC software	request data is not stored after processing, according to Apple	automatic routing or explicit PCC selection
ChatGPT integration without an account	OpenAI infrastructure	OpenAI	Apple and OpenAI state that requests are not stored for this integration and IP addresses are obscured	ordinary confirmation can be disabled in settings; files/photos need separate permission
ChatGPT with an account	OpenAI infrastructure	OpenAI	account settings and OpenAI policies apply	user deliberately signs in
Self-hosted MLX or llama.cpp model	the user’s Mac	user/model publisher	local unless the app adds telemetry or network tools	fully configurable by the user

Apple gives users control over ChatGPT request confirmations. Ordinary Siri confirmations can be disabled, but photos and files still require separate permission. The integration works without an account; signing in changes the governing data settings. Apple Mac support ³¹⁶

“Local” is not by itself a complete privacy audit. A local third-party app can log prompts, send analytics, or call remote tools. Conversely, PCC is cloud inference but uses a materially different security architecture from a typical hosted LLM API.

The comparison separates on-device processing, Private Cloud Compute, and conventional cloud AI by location, operator, retention, and user control.

Neural Engine, CPU, GPU, and unified memory

Apple silicon combines multiple compute engines. Their general roles are documented, but Apple does not disclose the exact operator placement for every internal Apple Intelligence model.

Component	Typical strength	Official platform role	Important caveat
CPU	control flow, tokenization, small operations, app/system logic	Core ML and Core AI can use the CPU	no public full operator map for Apple Intelligence
GPU	highly parallel matrix operations, local LLMs, image models	Metal, MLX, and Core ML can use GPU compute	actual use depends on runtime and model graph
Neural Engine	power-efficient supported ML operations	Core ML and Core AI can target the Neural Engine	not every LLM operator or framework automatically executes there
Unified memory	shared address space for CPU/GPU and other system components	MLX explicitly exploits Apple’s unified-memory architecture	shared does not mean unlimited; capacity and bandwidth remain constraints
NAND flash	persistent model storage	AFM 3 Core Advanced stores its full expert set in NAND	flash is not a substitute for active DRAM; selected experts are loaded into memory

Core ML can schedule supported workloads across CPU, GPU, and Neural Engine. Core AI, introduced in the 2026 platform generation, provides broader building blocks for generative AI. MLX uses unified memory so arrays do not require explicit copying between separate CPU and GPU memory pools. ^11121321

The stack shows how Apple Intelligence apps connect through the Foundation Models framework to models, compute engines, and memory; Apple's exact internal scheduling is not public.

How a local language model runs on Apple silicon

A simplified text-generation request involves:

1. tokenizing the input,
2. loading or activating model weights,
3. processing the prompt during the prefill phase,
4. building a KV cache for previous tokens,
5. generating output autoregressively,
6. detokenizing and passing the result through application and safety logic.

AFM 3 Core Advanced adds expert routing. The full expert library remains in flash, while selected experts are assembled into the active in-memory model for the prompt and can be periodically changed during generation. ⁴

Apple Foundation Models framework for developers

The Foundation Models framework is Apple’s native API for using the system model. In its 2025 form, it exposed the roughly 3B on-device model through Swift. An app does not need to ship the base weights, and the on-device route is designed to work offline. ⁸

Apple identifies good use cases such as:

• summarization,
• structured information extraction,
• text comprehension,
• rewriting and refinement,
• short-form dialogue,
• creative short content,
• app-specific workflows that call tools. ⁵⁸

The system model should not be treated as a universal source of current world knowledge. Narrow prompts, app-provided tools, retrieval from trusted local data, and schema-constrained output are safer design patterns.

Core API concepts

• LanguageModelSession: maintains state and coordinates requests.
• Guided Generation: returns typed Swift structures from a schema.
• Tool Calling: allows the model to invoke controlled application functions.
• Streaming snapshots: delivers progressively completed structured output.
• Availability and guardrails: lets an app check model, language, and restriction status.
• Adapters: version-specific LoRA adapters can specialize behavior but consume storage and require lifecycle management. ⁵⁸

2026 framework expansion

Apple’s WWDC 2026 material describes: ^9102223

• combined image and text input for the on-device system model,
• a broader LanguageModel and executor protocol,
• support for custom local and hosted model providers,
• planned Swift packages or integrations from vendors including Anthropic and Google,
• Dynamic Profiles,
• model and feature evaluation tools,
• an fm command-line tool and Python SDK on macOS 27,
• access to a PCC model for approved developers, subject to entitlements and quotas.

The result is closer to a model abstraction layer than a single local API. That flexibility creates a disclosure responsibility: applications should clearly indicate whether a request stays on-device, uses PCC, or goes to a third-party provider.

The diagram shows an app connecting through the Foundation Models framework to Apple's local system model, PCC, or an external provider protocol; see Apple's WWDC 2026 documentation.

Foundation Models framework vs Core ML, Core AI, MLX, and llama.cpp

Technology	Primary purpose	Model source	Typical compute	Can be offline?	Relationship to Apple Intelligence
Foundation Models framework	Apple system model and unified model API	Apple system model; 2026 PCC/providers	system-managed	yes for on-device route	direct
Core ML	deploy converted ML models in Apple apps	developer-provided	CPU, GPU, Neural Engine	yes	general ML layer, not the Apple Intelligence product
Core AI	native generative-AI and modern model pipelines	developer/platform	CPU, GPU, Neural Engine	model-dependent	complements system and custom model paths
MLX	research, fine-tuning, and flexible local inference on Apple silicon	community or self-trained	mainly CPU/GPU with unified memory	yes	separate from Apple Intelligence
llama.cpp	portable inference for quantized GGUF models	community	CPU/GPU with Metal support	yes	separate

Practical selection guide

• Choose the Foundation Models framework for private, narrow language tasks that benefit from a system-managed model and native Swift tooling.
• Choose Core ML when you need to ship or convert a specific model and control its app deployment; conversion typically runs through Core ML Tools.²⁴
• Choose Core AI for Apple-native generative pipelines and newer model abstractions.
• Choose MLX for Python-oriented research, fine-tuning, evaluation, and rapid local experimentation on a Mac.
• Choose llama.cpp for broad GGUF compatibility, portable local servers, and highly quantized community models.

Apple Intelligence vs a traditional local LLM on Mac

Criterion	Apple Intelligence	User-installed local LLM
Setup	OS-managed	user installs runtime and weights
OS context	deep, permission-controlled system integration	only data explicitly exposed by the app or user
Model choice	mostly controlled by Apple; some routes selectable	fully selectable
Customization	prompting, tools, typed output, limited adapters	system prompts, sampling, RAG, LoRA/fine-tuning, model swaps
Offline operation	available for many paths	available if no cloud tools are enabled
Cloud fallback	PCC or optional partner integration	only if the user configures one
Transparency	detailed papers, but closed production weights	varies by model card, training disclosure, and license
Maintenance	OS updates	user manages versions and storage

Apple Intelligence optimizes for low friction and operating-system integration. A self-hosted model optimizes for control, reproducibility, experimentation, and model choice.

Which one should you use?

If you want to…	Best fit	Why
summarize mail, rewrite text, clean up notifications	Apple Intelligence	built into the OS, no setup, mostly on-device
keep every prompt on your Mac, offline, with a model you choose	your own local LLM	no PCC or partner fallback unless you add one
ask long, knowledge-heavy questions or write code	your own 27B+ model or a cloud model	a ~3B system model is designed for focused tasks, not as a full chatbot
build a Swift app with on-device text features	Foundation Models framework	free system model, no weights to ship
fine-tune, use RAG over your files or compare models	MLX, llama.cpp or Ollama	full control over weights, context and data

Many Mac users end up with both: Apple Intelligence for system features and a local model in Ollama, LM Studio or MLX for private chat, documents and coding.

What your Mac needs for each path

Apple Intelligence 27 requires a supported Apple Silicon Mac with macOS 27. Disk space can be up to 8GB or 14GB depending on the device class; availability of older functions on earlier macOS releases is a separate question. Our own arithmetic from Apple’s 2025 report suggests the ~3B model’s 2-bit weights take well under 1GB of memory, but Apple does not publish the real RAM use per feature. ⁷⁵

A self-installed open model is a different story because you choose its size. The tiers below are illustrative capacity starting points, not measured token speeds or guarantees at 32K context. Context, quantization, KV-cache and macOS overhead must be checked for the specific model in the Mac LLM RAM calculator:

Your Mac’s unified memory	Indicative size tier for a quantized text model
8GB	Small 1B–4B models, with limited headroom beside macOS
16GB	Often 7B–12B using compact quantization and moderate context
24–32GB	Often 20B–30B at roughly Q4, depending on context
48GB	More headroom for 30B–35B and concurrent applications
64GB+	Entry point for strongly quantized 70B; throughput and longer context are not guaranteed

The actual memory calculation, including KV cache, is in the unified-memory guide. Apple Intelligence’s separate 8GB/14GB requirement refers to disk space, not this custom-model memory table.

Apple vs Google, Microsoft, and Samsung on-device AI

All major platform vendors now use hybrid architectures. The relevant differences are model management, developer APIs, hardware requirements, and cloud trust boundaries.

Platform	Local AI layer	Cloud/hybrid route	Developer surface	Distinguishing approach
Apple	Apple Foundation Models in the OS; Core ML/Core AI	Private Cloud Compute; optional partner models	Foundation Models framework, Core ML, Core AI, MLX	attested PCC architecture and deep native-app integration
Google/Android	Gemini Nano managed through AICore; ML Kit GenAI	Gemini cloud services and app APIs	ML Kit GenAI, LiteRT/LiteRT-LM, Gemini APIs	AICore manages supported on-device models and updates
Microsoft/Windows	local Windows AI components on Copilot+ PCs	Copilot and Azure services	Windows AI APIs, ONNX Runtime, DirectML	NPU-focused Copilot+ class with a 40+ TOPS platform threshold
Samsung Galaxy AI	mix of device-resident and cloud features	Samsung and partner services	primarily Galaxy/Android application surfaces	a device setting can prioritize on-device processing for supported features

Google documents Gemini Nano as an AICore-managed system service and exposes task-focused ML Kit GenAI APIs. Microsoft combines local NPU-backed Windows AI components with cloud Copilot experiences. Samsung distinguishes on-device and cloud-based Galaxy AI capabilities and offers a local-only preference on supported devices, with feature trade-offs. ^252627282930

No vendor is categorically “faster” or “more private” for every workload. A fair comparison must specify device, task, network state, model version, data path, and whether the result comes from vendor-run or independent evaluation.

Performance, storage, memory, and energy constraints

Officially documented facts

• Apple’s current requirement is up to 8GB or 14GB of disk storage depending on the supported Apple device; do not mistake that for RAM used during inference. ⁷
• The 2025 on-device model uses aggressive quantization and KV-cache optimization. ⁵
• AFM 3 Core Advanced stores full weights in NAND and activates 1–4B parameters for a request. ⁴
• Selected 2026 advanced capabilities require at least 12 GB of unified memory on specified iPad and Mac configurations. ¹⁸

Numbers Apple has not supplied

No reliable consumer table of system-model tokens per second by chip exists. Such a figure would depend on prompt length, output length, thermal state, active expert count, memory pressure, tool calls, model build, and decoding configuration.

The updated 8GB/14GB storage allowances are not simply one model’s weight-file size. It can cover multiple language, image, audio, safety, and support resources. Likewise, multiplying parameter count by quantization bits would not yield an exact download or RAM requirement: embeddings, metadata, adapters, runtime buffers, KV cache, and activations add overhead.

Chip-generation effects

Newer Apple silicon can improve memory bandwidth, GPU and Neural Engine throughput, power efficiency, and available unified-memory capacity. However, Apple grants advanced local features by supported-device list rather than publishing a universal performance formula. An M1 Mac remains compatible with Apple Intelligence but should not be assumed to run every AFM 3 Core Advanced workload announced for higher-tier 2026 hardware. ⁷¹⁸

Best use cases for Apple on-device AI

Strong local workloads

• summarizing short private documents or messages,
• extracting dates and structured fields,
• rewriting and tone adjustment,
• classification and prioritization,
• app-specific tool selection,
• compact assistants with bounded context,
• supported on-device speech and audio tasks.

Workloads more likely to need larger models

• complex multi-step reasoning,
• very long or richly multimodal contexts,
• large agentic tool chains,
• high-quality or photorealistic image generation,
• tasks requiring current web knowledge or broad external retrieval.

These are not absolute boundaries. A small model with constrained output and reliable tools can outperform a larger general model on a narrow workflow. A local execution path also does not eliminate hallucination or guarantee factual accuracy.

Advantages and limitations

Advantages

• low latency for suitable local tasks,
• offline operation,
• reduced data transfer,
• deep app and OS integration,
• system-managed weights without inflating every app bundle,
• native Swift APIs with typed output,
• a formalized privacy architecture for server fallback through PCC.

Limitations

• device, language, and regional fragmentation,
• closed production weights,
• limited world knowledge in the compact model,
• incomplete visibility into automatic routing,
• some features require networking or newer hardware,
• model and adapter behavior can change with OS releases,
• external integrations use separate policies and trust boundaries,
• Apple’s first-party evaluations are not substitutes for independent benchmarks.

Release status: September–October 2026

The iOS/iPadOS/macOS 27 software is available. Apple rolled out the major 2026 Apple Intelligence update on September 14, while Siri AI debuted as an English-language beta with additional languages and Health features scheduled in subsequent phases. Treating the entire platform as an unreleased preview is no longer accurate.

The new PCC platform is not Apple-silicon-only. Apple’s June 8 security disclosure documents deployment on Google Cloud using NVIDIA Confidential Computing GPUs, Intel TDX CPUs and Google Titan alongside the older Apple silicon PCC fleet. Apple maintains that all deployed PCC software must be Apple-approved and publicly verifiable. The June description said the Google Cloud privacy stack would ramp during the summer preview; our evaluation does not independently audit each live data-center node or its exact traffic share.

Some cloud-backed features have daily usage limits. The September release says expanded access may be offered for a fee, but provides no universal per-user daily count. For app developers, distinguish the on-device Foundation Models API, permitted PCC access and separately consented third-party services. Apple’s published routing policy remains high-level; a macOS screenshot is not evidence of which backend handled any one request.

The future of local AI on Apple devices

Apple’s 2026 architecture points to three longer-term developments:

• larger local models without full DRAM residency: prompt-level expert selection turns NAND into a model-capacity tier while keeping the active working set smaller;
• a common developer abstraction: Apple’s system model, PCC, and third-party models can be exposed through compatible protocols;
• more confidential cloud capacity: PCC expands from homogeneous Apple-silicon servers toward heterogeneous hardware and partner data centers.

The strategic question is no longer simply “local or cloud?” It is: which data reaches which model, under which verifiable trust boundary, for what purpose? Developers will need to treat data minimization, route disclosure, evaluation, fallback behavior, and permission design as core product features.

Conclusion: Hybrid on-device models plus Private Cloud Compute

Apple Intelligence is a hybrid on-device and private-cloud system, not one local LLM. Apple combines compact and increasingly capable device models with Private Cloud Compute and optional external providers. The progression from a quantized 3B model to AFM 3 Core Advanced demonstrates a practical route to larger local AI through sparse experts, NAND-backed capacity, and hardware-tiered activation. ⁴⁵

The most useful question is not merely whether a feature is “local.” It is: which model handles which data, where does it execute, and what storage, permission, and verification rules apply? Apple provides unusually detailed research and PCC security documentation. Even so, its complete feature-routing matrix, per-device memory use, and real-world model performance remain only partly disclosed.

How much RAM local models need and which runtime fits is covered in the unified memory guide and the Ollama setup.

Sources

Source review: October 13, 2026. The 27-series operating systems shipped on September 14, 2026; Siri AI and other capabilities still vary by region, device, language and beta availability. Architecture descriptions distinguish published Apple evidence from undisclosed production-routing details.

What remains unknown

• Apple’s exact automatic routing thresholds between on-device, PCC, and external providers are not public.
• A permanent matrix assigning every user-facing feature to one production model has not been published.
• Exact RAM use, individual model download size, and tokens per second by device are not disclosed.
• The complete operator placement of internal models across CPU, GPU, and Neural Engine is not documented.
• The 32K context cited in a WWDC 2026 session applies to the described developer PCC route and was not generalized to all system models.
• Feature availability and performance may change with later updates; iOS 27 and macOS 27 were publicly released September 14, 2026.
• No numerical thermal or energy comparison by Apple chip was included because Apple has not published a directly comparable official dataset.
• Rumors about unreleased chips, model sizes, or dates were excluded.

Footnotes

1. Apple Intelligence brings powerful AI capabilities into everyday experiences. Apple Newsroom, June 8, 2026. Press release. Supports operating-system integration and expanded system feature scope. https://www.apple.com/newsroom/2026/06/apple-intelligence-brings-powerful-ai-capabilities-into-everyday-experiences/ ↩ ↩² ↩³ ↩⁴

2. Private Cloud Compute: A new frontier for AI privacy in the cloud. Apple Security Research, June 10, 2024. Security architecture. Supports PCC requirements, attestation, data processing, and transparency. https://security.apple.com/blog/private-cloud-compute/ ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

3. Use ChatGPT with Apple Intelligence on iPhone. Apple Support, continuously updated. User documentation. Supports consent, photo/file confirmation, and account behavior. https://support.apple.com/guide/iphone/use-chatgpt-with-apple-intelligence-iph00fd3c8c2/ios ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶

4. Introducing the Third Generation of Apple’s Foundation Models. Apple Machine Learning Research, June 8, 2026. Official architecture/research article. Supports AFM 3 model names, parameter counts, sparse routing, NAND/DRAM design, training, and preview status. https://machinelearning.apple.com/research/introducing-third-generation-of-apple-foundation-models ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹

5. Apple Intelligence Foundation Language Models: Tech Report 2025. Apple, revised August 27, 2025. Technical report/paper. Supports the 3B on-device model, quantization, KV-cache sharing, model use cases, and adapters. https://machinelearning.apple.com/papers/apple_intelligence_foundation_language_models_tech_report_2025.pdf ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸ ↩⁹ ↩¹⁰ ↩¹¹ ↩¹²

6. Expanding Private Cloud Compute. Apple Security Research, June 8, 2026. Security architecture/preview. Supports Google Cloud and NVIDIA GPU expansion and the revised chain of trust. https://security.apple.com/blog/expanding-pcc/ ↩ ↩²

7. How to get Apple Intelligence. Apple Support, published September 14, 2026. Support documentation. Supports the current device, OS, 8GB/14GB storage, language, region, and feature-availability claims. https://support.apple.com/en-us/121115 ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

8. Meet the Foundation Models framework. Apple Developer, WWDC 2025. Technical session. Supports the on-device system model, Swift API, Guided Generation, Tool Calling, and offline operation. https://developer.apple.com/videos/play/wwdc2025/286/ ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷

9. WWDC26 Apple Intelligence guide. Apple Developer, June 2026. Developer overview. Supports 2026 framework, Image Playground, and platform changes. https://developer.apple.com/wwdc26/guides/apple-intelligence/ ↩ ↩² ↩³ ↩⁴

10. Bring an LLM provider to the Foundation Models framework. Apple Developer, WWDC 2026. Technical session. Supports the LanguageModel/executor protocol, third-party routes, and the developer PCC context statement. https://developer.apple.com/videos/play/wwdc2026/339/ ↩ ↩² ↩³

11. Core ML. Apple Developer Documentation, continuously updated. Framework documentation. Supports CPU, GPU, and Neural Engine execution. https://developer.apple.com/documentation/coreml/ ↩ ↩²

12. Core AI. Apple Developer Documentation, 2026. Framework documentation. Supports Apple’s generative-AI and model-pipeline platform. https://developer.apple.com/documentation/coreai ↩ ↩²

13. MLX. Apple Machine Learning Research, GitHub, continuously updated. Official open-source repository. Supports MLX’s purpose and Apple-silicon optimization. https://github.com/ml-explore/mlx ↩ ↩² ↩³

14. llama.cpp. ggml-org, GitHub, continuously updated. Official open-source repository. Supports local GGUF inference and Metal acceleration on Apple silicon. https://github.com/ggml-org/llama.cpp ↩ ↩²

15. Introducing Apple Intelligence, the personal intelligence system. Apple Newsroom, June 10, 2024. Press release. Supports the original product definition and hybrid architecture. https://www.apple.com/newsroom/2024/06/introducing-apple-intelligence-for-iphone-ipad-and-mac/ ↩ ↩²

16. OpenAI and Apple announce partnership. OpenAI, June 10, 2024. Official provider announcement. Supports privacy handling for the no-account Apple integration. https://openai.com/index/openai-and-apple-announce-partnership/ ↩ ↩²

17. Introducing Apple’s On-Device and Server Foundation Models. Apple Machine Learning Research, June 10, 2024. Official research article. Supports the first publicly described Apple Intelligence model generation and specialization. https://machinelearning.apple.com/research/introducing-apple-foundation-models ↩

18. Apple Intelligence and Siri. Apple, continuously updated product page, July 2026 snapshot. Product and availability documentation. Supports 2026 feature announcements, device footnotes, and preview status. https://www.apple.com/apple-intelligence/ ↩ ↩² ↩³ ↩⁴ ↩⁵

19. Apple Intelligence gets even more powerful with new capabilities across Apple devices. Apple Newsroom, June 9, 2025. Press release. Supports Live Translation, developer access, and 2025 platform features. https://www.apple.com/newsroom/2025/06/apple-intelligence-gets-even-more-powerful-with-new-capabilities-across-apple-devices/ ↩

20. Use Apple Intelligence models in Shortcuts on iPhone. Apple Support, continuously updated. User documentation. Supports explicit selection of On-Device, PCC, or ChatGPT in Shortcuts. https://support.apple.com/guide/iphone/use-apple-intelligence-in-shortcuts-iph78c41eaf8/ios ↩

21. Unified Memory in MLX. Apple MLX Documentation, continuously updated. Technical documentation. Supports the shared CPU/GPU memory model used by MLX. https://ml-explore.github.io/mlx/build/html/usage/unified_memory.html ↩

22. Use the Foundation Models framework from the command line. Apple Developer, WWDC 2026. Technical session. Supports the fm CLI and Python SDK on macOS 27. https://developer.apple.com/videos/play/wwdc2026/334/ ↩

23. Access the cloud model with the Foundation Models framework. Apple Developer, WWDC 2026. Technical session. Supports developer PCC access, eligibility, and quotas. https://developer.apple.com/videos/play/wwdc2026/319/ ↩

24. Core ML Tools overview. Apple, continuously updated. Developer documentation. Supports model conversion and ML Program deployment. https://apple.github.io/coremltools/docs-guides/source/overview-coremltools.html ↩

25. Gemini Nano. Android Developers, continuously updated. Platform documentation. Supports AICore-managed on-device Gemini Nano. https://developer.android.com/ai/gemini-nano ↩

26. AI on Android. Android Developers, continuously updated. Platform overview. Supports local and cloud AI developer routes. https://developer.android.com/ai/overview ↩

27. NPU devices / Copilot+ PCs. Microsoft Learn, continuously updated. Platform documentation. Supports the Copilot+ NPU class and requirements. https://learn.microsoft.com/en-us/windows/ai/npu-devices/ ↩

28. Windows Copilot AI components. Microsoft Support, continuously updated. Product support. Supports local Windows AI components. https://support.microsoft.com/topic/windows-copilot-ai-components-a9ef14e9-32a7-497f-b780-7b6fb63af793 ↩

29. Galaxy AI. Samsung, continuously updated product page. Product documentation. Supports the Galaxy AI feature family. https://www.samsung.com/us/galaxy-ai/ ↩

30. On-device and cloud-based Galaxy AI functions. Samsung Support, continuously updated. Support documentation. Supports the local/cloud distinction and device-processing preference. https://www.samsung.com/ae/support/mobile-devices/how-to-distinguish-between-on-device-functions-and-cloud-based-functions-among-new-ai-features/ ↩