Apple Intelligence uses a hybrid on-device and cloud approach called Private Cloud Compute, while open source tools like Cloaked run entirely on-device with no cloud processing ever. Apple offers tight iOS integration with no setup required; Cloaked offers 15+ open source model choices and works on iPhone 12 and later.
Both represent a genuine step forward from sending your conversations to a conventional cloud AI. But they make different trade-offs, and understanding those trade-offs helps you choose the right tool for what you actually care about.
What Apple Intelligence Actually Does
Apple Intelligence is Apple’s umbrella term for the AI features built into iOS 18, iPadOS 18, and macOS Sequoia. It covers a broad range of capabilities: writing tools, image generation, notification summaries, Siri improvements, and a built-in chatbot experience that can optionally tap into ChatGPT.
What Apple Intelligence is not, despite the branding, is a purely on-device system. Apple uses what it calls Private Cloud Compute — a purpose-built cloud infrastructure designed to handle requests that require more compute than the device can provide. When a request is too demanding to process locally, it is routed to Apple’s servers.
Apple has been transparent about this architecture and has made strong technical commitments around it: requests are not logged, Apple employees cannot inspect them, and the server software can be independently audited. These are meaningful protections that go well beyond what most cloud AI providers offer.
Apple Intelligence launched in 2024 and is available on iPhone 15 Pro, iPhone 16, and later models, as well as Macs and iPads with M-series chips — representing a hardware requirement that excludes hundreds of millions of older Apple devices.
The distinction between “what processes on-device” and “what goes to Private Cloud Compute” is not always visible to users. Simpler tasks tend to stay local; complex writing, summarization, and chatbot queries are more likely to involve the cloud. Apple has not published a complete list of which features go where, which makes it difficult to predict the data path for any given request.
What Open Source On-Device AI Does Differently
The alternative approach — used by apps like Cloaked — is to run inference entirely on your device, for every request, with no cloud fallback.
This is possible because of recent advances in model efficiency. Quantized open source models from Meta, Google, Microsoft, Alibaba, and others now fit comfortably in the memory of modern iPhones. A 4-bit quantized 7B model runs at reasonable speeds on an A16 or newer chip. Smaller models like Llama 3.2 3B and Phi-4 Mini run well on even older hardware.
Cloaked uses the Apple MLX framework — specifically MLX Swift — to handle inference. Every token of every response is generated by the chip in your hand. There is no network call, no server, no cloud infrastructure of any kind involved in the conversation.
This means the privacy guarantee is architectural rather than policy-based. The data never leaves your device because there is no mechanism by which it could. A privacy policy can change; an architecture that lacks a network path cannot change what it is.
Cloaked supports 15+ open source models, ranging from 400MB (Qwen 2.5 0.5B) to approximately 4GB (Mistral 7B), and works on iPhone 12 and later running iOS 17.
For a deeper look at the technical architecture behind this approach, the complete guide to on-device AI covers how local inference works and what it requires.
Hardware Requirements: A Meaningful Difference
One of the most practical differences between the two approaches is which devices they support.
Apple Intelligence requires iPhone 15 Pro or iPhone 16 and later (or M-series Macs and iPads). The feature simply does not appear on older hardware, including the standard iPhone 15, iPhone 14 series, and anything before that. Apple’s stated reason is that on-device processing for Intelligence features requires the hardware capabilities introduced with the A17 Pro chip.
Cloaked works on iPhone 12 and later, running iOS 17. That covers five additional generations of hardware. For users with iPhone 12, 13, 14, or the standard iPhone 15, open source on-device AI is the only path to genuinely private local inference. The performance scales with the chip — an A18 Pro will run larger models faster — but even older supported devices run the smaller models at usable speeds.
This matters practically: a large share of active iPhones in use today are models that Apple Intelligence does not support. Open source tools fill that gap.
Model Choice and Transparency
Apple Intelligence uses Apple’s own proprietary models. The architecture is not public, the weights are not downloadable, and independent researchers cannot audit what the model has been trained on or how it behaves. Apple provides documentation of the system design and security properties, but the models themselves are a black box.
Open source models are different in kind. Models like Llama 3.2 (Meta), Gemma 3 (Google), Phi-4 Mini (Microsoft), and Mistral 7B come with published model cards, training data disclosures, and — in many cases — permissive licenses that allow independent evaluation and modification. Researchers can and do test these models for biases, capabilities, and failure modes. The weights are publicly available.
As of early 2026, the Hugging Face model hub hosts over 1.2 million publicly accessible model repositories, with open source language models accounting for a significant and growing share of that catalog.
Beyond auditability, open source models give users meaningful choice. Different models have different strengths: Phi-4 Mini handles reasoning and coding well, Gemma 3 4B is a capable general-purpose model, DeepSeek R1 1.5B is optimized for step-by-step reasoning, Qwen 2.5 supports strong multilingual performance. Being able to select the right model for the right task is something Apple Intelligence does not currently offer.
For a detailed breakdown of how these models compare, the guide to open source AI models covers the trade-offs across the current generation of locally-runnable models.
Privacy Architecture: Similar Goals, Different Paths
Both Apple Intelligence and fully on-device open source AI are trying to solve the same problem: how do you give users powerful AI without requiring them to hand over intimate data to a cloud provider?
Apple’s answer is Private Cloud Compute — a hardened cloud infrastructure with strong technical controls, independent auditability, and no persistent logging. It is a genuine attempt to make cloud AI trustworthy, and the security properties Apple has documented are substantially better than what conventional cloud AI providers offer. Apple deserves credit for taking this seriously.
The fully on-device answer is architectural elimination. If inference never leaves the device, there is no cloud infrastructure to harden, no logging to disable, no audit process to trust. The threat model shrinks because the attack surface shrinks. A subpoena cannot extract conversations from a server that never received them. A data breach at an AI provider cannot expose conversations that were never transmitted.
Apple’s Private Cloud Compute security review found that requests processed in the cloud are not associated with Apple IDs, are not retained after the response is returned, and cannot be accessed by Apple employees — a substantially stronger set of commitments than those offered by most cloud AI services.
The trade-off is capability. Private Cloud Compute lets Apple Intelligence handle requests that would be too slow or too demanding for on-device processing. Fully on-device AI is bounded by local hardware — which is increasingly capable, but not unlimited. Users with complex tasks that strain local models will sometimes notice the difference.
Offline Use
Apple Intelligence requires an internet connection for any request that routes to Private Cloud Compute. On-device features work offline, but it is not always clear in advance which path a given request will take.
Cloaked works fully offline after the initial model download. Once a model is on your device, every conversation happens locally regardless of network state. This matters in contexts where connectivity is limited or untrusted: on flights, in locations with spotty coverage, or in situations where you do not want your traffic visible to a network observer even before it reaches a server.
The Bottom Line
Apple Intelligence and open source on-device AI are not competitors in the usual sense — they serve somewhat different use cases and make different bets about what users value most.
Apple Intelligence is the right choice if you want AI deeply integrated into your iOS workflows — writing suggestions in Mail, notification summaries, Siri that understands context across apps — and you are on supported hardware. Apple’s privacy engineering is serious, and Private Cloud Compute is meaningfully better than conventional cloud AI. The system just works, with no setup required.
Open source on-device AI is the right choice if you want a guarantee that your conversations are physically incapable of leaving your device, you want to choose between 15+ models based on your needs, you are on older hardware that Apple Intelligence does not support, or you want to use AI fully offline. You give up the deep iOS system integration; you gain architectural privacy and model flexibility.
The two are not mutually exclusive. Apple Intelligence handles system-level features; a fully on-device app handles sensitive conversations. Using both is a reasonable approach.
If you want the fully on-device option with no cloud processing, no accounts, and 15+ open source models to choose from, Cloaked is available on the App Store. It works on iPhone 12 and later. The first model download takes a few minutes. After that, every conversation stays on your device — by design, not by promise.