The Future of Apple Intelligence (and Why It's Worth Getting Fluent Now)

This is Part 6 of a 6-part series on Apple Intelligence. Start with Part 1: What Is Apple Intelligence? if you missed it.

Everything in the last post (Writing Tools, the Shortcuts “Use Model” action, the offline tap-to-AI loop) is the floor, not the ceiling. That’s the part I keep coming back to: what runs on the phone in your pocket today is the least capable version of on-device AI you will ever use.

So this last post is where I make my case for why that’s worth paying attention to now, and lay out where I think it goes. Fair warning up front: most of what follows is my opinion and prediction, not fact. I’ll hedge it, and I’ll give you the honest argument against it too.

The one dynamic behind all of this

The basic dynamic is simple. As the silicon gets better, the on-device model handles more of the work the cloud used to do. Go back to the assistant analogy from the earlier posts: the junior at the desk keeps getting sharper, so fewer requests need to walk down the hall to the expensive specialist.

Two things make that direction hard to ignore. On-device inference has no per-request bill (it’s roughly free once the model is on your phone), and it never leaves the device, so it’s private by default. The counterweight is capability, and that gap is closing. Some industry analysis argues small on-device models now cover a large share of the everyday jobs (summarizing, classifying, pulling structured data out of messy text) while the cloud still clearly wins the hard, multi-step reasoning. As that gap shrinks, the reason to pay the cloud tax shrinks with it.

That’s the whole setup. More work slides down to the desk over time. My predictions are all just guesses about how Apple rides that curve.

Prediction 1: capable models gated to premium hardware

Apple has already taken the first step here, which is part of why I’m confident about the rest. As of 2026 there isn’t a single on-device model anymore. There’s a baseline model (roughly 3 billion parameters) and a more capable 20-billion-parameter sparse model, AFM 3 Core Advanced, which Apple describes as unlocked by its most capable hardware. My prediction is that Apple leans hard into that tiering and turns a smarter local model into a standing reason to buy the newer, pricier device.

People already do a version of this by hand. If you run local models with something like Llama or Ollama, the size you can load is capped by your RAM (the memory-not-TOPS point from Part 3), so better hardware means a smarter local model. I think Apple keeps pushing that dynamic into its product tiers.

Reporting suggests Apple wants to go further still. A report from The Information (via MacRumors, July 2026), and I want to be clear this is reporting, not something Apple has confirmed, says Apple met with a startup called PrismML that squeezed a 27-billion-parameter model onto an iPhone 17 Pro, larger than even Apple’s 20-billion-parameter AFM 3 Core Advanced. The stated motive was exactly the dynamic above: run more features on the device to cut cloud cost and improve privacy. Note the demo hardware. It was the Pro.

There’s already precedent for the gating half of this. Apple Intelligence has always required newer hardware (an 8 GB RAM floor, specific chip generations), so tying capability to the device tier is not a new move for them. The honest complication is cost: bigger on-device models need more memory, which pushes the hardware price floor up. Analysts have separately estimated future Pro pricing could climb toward $1,399 as memory costs rise. That supports the “gated to premium” part of my prediction while cutting against any “everyone gets a smarter model” reading.

Prediction 2: a local inference box you plug into your setup

This one is speculation, and I want to flag it loudly. There is no reported Apple product here. It’s me extrapolating.

The guess: a small Apple appliance that acts as your own private, local inference server, so your Mac and phone offload heavier AI work to a box in your house instead of the cloud. Free, private, and more capable than what fits on a phone.

A few real signals make me think the pattern is at least plausible (none of them is proof):

macOS 27 introduced a command-line tool called fm that runs Apple’s models locally from the terminal (on-device by default, with a switch to Private Cloud Compute), shown at WWDC26. Third-party tooling has already wrapped it into an OpenAI-style local server.
WWDC26 had a whole session on running local agentic AI on the Mac using MLX, Apple’s on-device machine-learning framework.
Enthusiasts already press a Mac mini into service as a home AI server today.
Apple is reportedly building its own AI-inference server chip (codenamed “Baltra,” with Broadcom), though that’s for Apple’s own data centers, not a thing you’d buy.

So the building blocks exist. Whether Apple ever turns them into a consumer product is genuinely a coin flip, and I’d put lower confidence on this than on the first prediction.

The honest case against me

I’d be selling you a one-sided story if I stopped there, because Apple is betting big on the cloud at the exact same time.

The marquee 2026 feature, the new Siri, goes the cloud route. Reporting from Bloomberg’s Mark Gurman puts it at roughly $1 billion a year for a custom 1.2-trillion-parameter model running Siri’s cloud reasoning on Private Cloud Compute. Apple’s own language is drier, describing a top server model built in collaboration with Google, and I’d treat the dollar figure and the parameter count as reporting, not confirmed fact.

That’s not the only cloud signal. In 2026 Apple officially expanded Private Cloud Compute beyond its own data centers onto Google Cloud (running on NVIDIA GPUs) to serve larger models, and it framed this as an expansion running in parallel with its Apple-silicon servers, not a migration off them. So when I say the local tier is eating the cloud’s job, understand that Apple is also scaling the cloud up to chase raw capability. The near-term reality looks hybrid: each request gets routed to wherever it’s cheapest and most private to run well.

The wild-West question, answered

Back in Part 1 I floated an opinion I called my “wild West” read, and this is where I close it out. My read (and I can’t prove this) is that the field right now optimizes for performance and new use cases more than for cost or privacy. Given the capability jumps of the last two years, that’s a reasonable place for the industry to spend its attention. But I expect the priorities to shift toward cost and privacy over the next few years.

The grounding, without overclaiming it: on-device inference is roughly free at the margin, dozens of countries now have data-localization rules that make keeping data on the device easier to comply with, and the capability gap keeps narrowing. Apple’s own reported reason for chasing PrismML (do more on-device to cut cost and boost privacy) is Apple already moving in that direction.

The honest version of my prediction: more local where local is good enough, on hardware that keeps getting better, while the cloud keeps doing the genuinely hard stuff. The 1.2-trillion-parameter cloud model is all the proof you need that the performance-first era is still very much running.

Why get fluent now

Which brings me to why I’d bother learning any of this today, before the future arrives.

For anyone using an iPhone: the new Siri plus App Intents (the plumbing that lets it take actions across your apps, from Part 2) is starting to look more agent-like, something you delegate tasks to rather than issue commands at. The people who get comfortable with the on-device building blocks now (the Shortcuts “Use Model” action, offline automations) will be the ones ready to actually use the agent layer when it fully lands.

For developers, the pitch is sharper. The Foundation Models framework gives you free, private, offline inference on the same model that ships in the OS. You can build AI features with zero cloud cost and a lot less data-handling liability, in a few lines of Swift. If you want the mental model, Apple’s WWDC session on the framework is the place to start. Getting that mental model now is a head start on where I think the whole thing is heading.

That’s the series

I went into this expecting Apple Intelligence to be a gimmick, and came out convinced the local-first bet is the genuinely interesting part of AI right now. The most interesting part, to me, is the quiet one: a small model on the device in your pocket, free and private and offline, that can do more of the work every year.

If you jumped in here, the whole thing starts at Part 1, and the five posts in between cover the history, the models, the privacy guarantees, and the things you can actually do today. If you want to catch whatever I write next, the site has an RSS feed you can point your reader at. Thanks for reading along this week.

Frequently asked questions

Will Apple release bigger AI models only on newer iPhones? It’s already started. In 2026 Apple added a more capable 20-billion-parameter on-device model (AFM 3 Core Advanced) above the baseline 3-billion-parameter one, tied to its most capable hardware. My prediction (an educated guess, not a confirmed plan) is that Apple leans harder into that tiering and makes a smarter local model a regular reason to buy a higher-end device. Reporting that it explored even larger on-device models on Pro-tier silicon points the same way.

Is Apple moving away from the cloud? No. Even as more work runs on-device, Apple is scaling its cloud up: it officially expanded Private Cloud Compute onto Google Cloud in 2026, and reporting says it’s paying for a very large custom model to power the new Siri’s cloud reasoning. The realistic near-term picture is hybrid, not one or the other.

Is the standalone AI box a real Apple product? No. That’s my speculation. The pieces exist (a local fm command-line tool in macOS 27, on-device MLX agents, people running Mac minis as home AI servers), but Apple has not announced any such consumer device.

Do developers have to pay to use the on-device model? No. The Foundation Models framework gives on-device inference that’s free, offline, and has no per-token cost. Access to the Private Cloud Compute tier is also offered free to smaller developers.

What’s the fastest way to start using Apple Intelligence today? The Shortcuts “Use Model” action, covered in Part 5. It lets you send a prompt to the on-device model (or Private Cloud Compute, or ChatGPT) from a shortcut you build yourself, no code required.