Buy an iPhone 16 today and Apple's on-device model has roughly 3 billion parameters. Buy an iPhone 17 Pro and it can have close to 20 billion — same operating system, same Apple Intelligence brand, a different model quietly selected by which silicon is under the glass. That gap is the story of on-device AI in 2026: the marketing is unified, the hardware is not, and the distance between the two is exactly what a builder needs to map before writing a line of code.

What Apple actually shipped

Apple's developer-facing move came at WWDC 2025, when it opened its on-device model to third parties. The new Foundation Models framework gives developers direct access to the on-device foundation language model at the core of Apple Intelligence, built around a roughly 3-billion-parameter model. All of it runs on-device, so data going into and out of the model stays private, and it can run offline. Apple was explicit about the model's limits: it excels at text tasks like summarization, entity extraction, and short dialog, but it is not designed to be a chatbot for general world knowledge.

That framework shipped to real apps quickly — journaling app Day One, made by Automattic, used it to add privacy-friendly intelligence, with the company saying it helped them rethink what's possible in the app. But it arrived with a hard ceiling that matters for anyone planning a feature: one developer account of a WWDC 2026 lab session describes the on-device model's context window as unchanged, a shared input-and-output budget that leaves little room for long conversations once a prompt gets long.

A year later, at WWDC 2026, Apple raised that ceiling for newer hardware only. AFM 3 Core Advanced, Apple's most powerful on-device model, is natively multimodal and uses a 20-billion-parameter sparse architecture that activates just 1 to 4 billion parameters at a time, unlocked by and optimized for Apple's most capable silicon. Older Apple Intelligence phones don't get left with nothing — the family also includes AFM 3 Core, a smaller 3-billion-parameter dense model that runs more broadly — but the standard 8GB memory requirement that has applied since Apple Intelligence launched no longer guarantees the most capable on-device features, marking the first time Apple raised the bar for them. MacRumors reported the 20B model specifically requires an iPhone 17 Pro or iPhone Air.

The more consequential change for builders may be a quieter one. Until WWDC 2026, the Foundation Models framework had one rule — Apple's on-device model or nothing — but that changed with a public protocol layer that any LLM, cloud API, or self-hosted model can implement. That turns Apple's session-and-tool-calling plumbing into a distribution layer developers can point at other models, not just Apple's own.

What Google actually shipped

Google's path runs through a system service called AICore. Gemini Nano runs in Android's AICore system service, which leverages device hardware for low inference latency, and ML Kit's GenAI APIs are built on top of it to enable on-device execution of GenAI foundation models. Unlike Apple's single general-purpose session API, Google shipped a set of named, task-specific endpoints: Prompt for custom text or multimodal generation, Summarization for articles or conversations, Proofreading for short chat messages, Rewriting in different tones, Image Description, and Speech Recognition to transcribe audio.

Device support is unusually specific and cross-vendor. GenAI APIs are available on Android devices using optimized MediaTek Dimensity, Qualcomm Snapdragon, and Google Tensor platforms through AICore. Google's own documentation lists supported hardware down to the model number, including Pixel 9 through 11 series phones alongside specific Honor, Oppo, Poco, and Samsung Galaxy S25 devices — a rare case of a vendor publishing a checkable compatibility list rather than a blanket claim.

On Pixel hardware specifically, the feature history is well documented and mostly genuine on-device processing: Pixel 9's Call Notes and Pixel Screenshots shipped in August 2024, multimodal Gemini Nano arrived in November 2024, Scam Detection for calls and messages in March 2025, and Pixel 10 with Tensor G5 added Nano-powered Magic Cue and Voice Translate in October 2025. Worth flagging for builders comparing platforms: not every AI feature that looks similar runs the same way. Android Authority's device testing found Samsung's Voice Recorder app includes a similar AI summarization feature, but it relies on Google's larger Gemini Pro model that can only run in the cloud, trading privacy and speed for the ability to handle hours-long recordings. Same category of feature, different architecture entirely — exactly the kind of detail that only shows up when you check the implementation rather than the feature list.

What Samsung actually shipped

Samsung has been building Galaxy AI into its phones since the Galaxy S24 series, and the Galaxy S26 Ultra marks a shift in where it's headed — from reactive tools toward proactive suggestions. The phone now learns habits and surfaces suggestions before being asked — a text about a meeting might prompt the phone to suggest opening the calendar, or a friend's photo request can pull up relevant gallery images automatically. Samsung attributes this to hardware: the faster NPU in the Snapdragon 8 Elite Gen 5 for Galaxy allows more of these tasks to run directly on the device rather than relying on a cloud connection.

The specific performance numbers, though, are vendor claims rather than independent measurements. Samsung says the phone's custom Snapdragon 8 Elite Gen 5 for Galaxy delivers up to 19% better CPU performance, 39% better NPU performance, and 24% better GPU performance than the prior generation — and the reviewer covering it was careful to note those are Samsung's numbers, not independent guarantees. One feature illustrates that hardware capability and shipped availability are separate questions entirely: reporting on Samsung's community forums found Call Screening is primarily a software feature that runs directly on the device, meaning phones like the Galaxy S25 likely have enough power and NPU capacity to handle it — yet Samsung confirmed support starts only with the Galaxy S26 series. Whether that's an engineering constraint or a product-segmentation decision, Samsung hasn't said publicly. Either way, the lesson holds: NPU capability on a spec sheet doesn't guarantee a feature actually ships to your device.

The pattern underneath all three

Line up the three platforms and a common structure appears. Each vendor now offers a genuinely local inference path — Apple's Foundation Models framework, Google's AICore plus ML Kit GenAI APIs, Samsung's Galaxy AI features running on-device where NPU and memory allow. Each also reserves a second, larger tier of capability for the cloud, invoked automatically when the task exceeds what the local model can do. Apple calls this Private Cloud Compute; Samsung uses Gemini Pro in the background for some Galaxy AI features without necessarily flagging the handoff to the user. The practical result for a builder is that on-device AI is not a single guarantee — it is a spectrum that shifts by device, region, and even which specific feature inside an app you're calling.

What to check before you build

A few concrete, verifiable things are worth confirming directly against current documentation rather than a launch keynote: which of your target devices actually ship the advanced model tier versus the base one, since Apple's 20B model is silicon-gated, not just OS-version-gated; the real context-window budget for the framework you're using, since a shared input-output token limit changes what summarization can mean in practice; whether a feature you're evaluating actually completes on-device or silently falls back to a cloud model when inputs get long, as documented in Samsung's own Voice Recorder implementation; and Google's published device-support list for ML Kit's GenAI APIs, which names exact models rather than gesturing at a chipset family. None of this requires trusting a press release — all of it is checkable in developer documentation before a single feature ships.