Are the New iPhone 18 Pro / Pro Max / Duo Really 2x for AI? What Matters If You Run Your Own Local Model

2026年9月21日

Apple says the new iPhone has “2x AI performance." Running a local LLM on my own machine, I’m curious whether that 2x is the kind of 2x that actually matters for how I use it.

Three phones carry the new chip this time. Does that 2x make the model I’m already running go any faster?

What This Article Covers, and How I Checked It

Three things: where the new chip’s “2x" actually matters, what changes when running an AI model you chose yourself, and whether the three phones differ from each other at all.

I own none of these phones. There is not a single number here measured on real hardware. I’ve split this into what Apple’s official page states, and what comes from various write-ups where the official page says nothing.

What’s Officially Stated

The following table is read directly from Apple’s announcement materials and product page.

Item	Apple’s stated figure
Chip name	A20 Pro
AI circuitry	Two 16-core Neural Engines, 32 cores combined. 2x the AI performance of the previous generation (A19 Pro)
Memory bandwidth (how fast data moves)	50% wider than the previous generation
Amount of memory	Not stated
Running third-party models	Not stated

The amount of memory isn’t stated. For local LLMs, that’s the first thing that decides whether something even runs. Since Apple hasn’t published it, I can’t state it either.

Where Does the “2x" Actually Matter?

This is about the Neural Engine doubling. There’s one catch with this 2x.

The main route to the Neural Engine is Apple’s own framework, Core ML. Where a given operation actually runs is decided by the OS, so going through Core ML doesn’t guarantee the Neural Engine either. Meanwhile, llama.cpp — the software commonly used to run GGUF-format models locally — doesn’t go through Core ML in its usual build. It uses the GPU and CPU instead.

In other words, for anyone choosing and running their own AI model, the 32-core Neural Engine doesn’t come into play as-is.

So what does change? Memory bandwidth, which is now 1.5x. Apple writes it as “50% wider," which is the same thing. Bandwidth is how fast data moves to and from memory.

From what I’ve measured so far, write speed is roughly determined by this division:

Write speed ≈ memory bandwidth ÷ amount read per token

If the numerator grows to 1.5x, the same AI model gets straightforwardly faster. For anyone running an AI model they chose, that 1.5x is more relevant than the “2x AI performance" headline.

I looked into the same shape of claim on the Mac before. The Mac Studio’s “4.3x AI performance" traced through the Neural Engine and MLX. Same structure here.

“20B Runs on an iPhone" — Which AI Model Is That?

There’s talk of a large AI model running on-device. That 20B figure refers to an AI model Apple built itself.

Various write-ups state the following. None of the four points below could be confirmed on the official page.

• What runs is Apple’s own AI model. You cannot swap in an AI model of your choice
• You also can’t adjust the weights yourself
• Running it is reportedly said to require 12GB of memory on the previous generation
• Large weights are kept in storage, with only a portion loaded for each individual response

For anyone who wants to choose their own AI model, this falls outside the scope entirely. It’s a powerful setup if you’re using what Apple provides, but that’s a different story from “run whatever AI model you like on your phone."

Can You Ship a Small AI Model Inside an App?

Apple’s own AI model can’t be swapped out. So is there no route to running an AI model of your choosing? There is one, if you put a small AI model inside an app. llama.cpp’s official repository carries an iOS example (llama.swiftui).

The wall here isn’t only the phone’s total memory. iOS terminates apps that use too much of it. Having 12GB in the device doesn’t mean one app can use all of it.

Various write-ups put the realistic line as follows. These are not Apple’s figures.

Item	Reported guideline
AI model size	Up to the 2-billion-parameter class, around 1–2GB at 4-bit
Quantization	Q4_0 or Q4_K_M; anything larger leaves no headroom
What pushes you over	Context length matters more than the weights
Write speed	15–30 tokens per second on A17 Pro and later
Heat	Reported to drop about 44% within minutes of sustained use

It does end up simpler than Apple’s. Against Apple’s 20 billion, what you can ship yourself is the 2-billion class — a full order of magnitude apart.

And here’s where Apple’s placement tells. Apple’s AI model keeps its weights outside your app, so they don’t land on your app’s own usage (which is not the same as saying the device uses no memory for it). Carry an AI model yourself and the whole thing comes out of your budget. On the same phone, the two sit in different places.

The route is still open. If being able to pick the AI model yourself is what matters, shipping a 2-billion-class one is a workable choice. The more memory a phone has, the more room there is for it.

How Much Memory Is There? (This Part Is a Guess)

Various write-ups estimate the following. Apple has not disclosed this.

Model	Estimated memory
iPhone 18 Pro / 18 Pro Max / iPhone Duo	12GB
iPhone 18 / 18e	9GB

There’s a clue for these numbers: the 17 series. Those figures surfaced through developer tooling, and the feature split had already happened there.

Model	Memory	Top-tier on-device AI in iOS 27
iPhone 17 (base)	8GB	Not available
iPhone 17 Air	12GB	Available
iPhone 17 Pro / Pro Max	12GB	Available

That’s the turning point. Until now, the condition for Apple’s AI features was a single line at “8GB or more." With iOS 27 it split into what runs on 8GB and what needs 12GB. What the 8GB iPhone 17 (base) lost was reportedly the top-tier on-device AI model and Siri voice customization. Most features — photo and mail search among them — still run on 8GB.

For that top-tier AI specifically, iPad is said to require M4 or later with 12GB or more, and Mac M3 or later with 12GB or more (a separate matter from the conditions for Apple’s AI features as a whole). The number 12GB keeps turning up across product lines.

That line matters for anyone running an AI model of their own too. Apple itself has drawn a boundary saying some AI models won’t fit under 12GB. Memory in the device now decides directly whether a feature exists.

As for the iPhone 18 (base), which isn’t out yet, the estimates disagree. Some say 9GB, others say it climbs to 12GB. At 9GB, the two features above are reportedly still unavailable. This isn’t a matter of single gigabytes — there’s a line at 12GB.

Even assuming 12GB, not all of it is usable. Some goes to display, and the system itself is running too. On my own PC, a 16GB graphics card only had 15.4GiB actually usable.

If the drop-off is roughly proportional, out of 12GB, maybe the low 8GB range is available for the AI model. This estimate is my own guess — I have not confirmed it. At that scale, I’d guess an AI model in the 4–8 billion parameter range, cut down to 4-bit, is the realistic ceiling.

Do the Three Phones Change the Conditions?

Here are the three side by side, read directly from Apple’s official specs pages.

Item	iPhone 18 Pro	iPhone 18 Pro Max	iPhone Duo
Chip	A20 Pro (identical across all three)
AI circuitry	Two 16-core Neural Engines (identical across all three)
Amount of memory	Not stated (for any of the three)
Display	6.3 inches	6.9 inches	7.6 inches open / 5.4 closed
Storage	256GB–2TB (all three)
Battery (video playback)	Up to 36 hours	Up to 45 hours	Up to 31 hours inner / 44 outer

Various write-ups put the Duo’s price starting at ¥364,800, releasing October 23. Other points mentioned include matching aspect ratios between the inner and outer screens, fingerprint authentication on a side button, and Apple Pencil support.

From a local-LLM standpoint, all three are the same. The chip and the AI circuitry are identical, and the amount of memory is unstated for every one of them. I found nothing that would make one of them run faster than another.

What differs is display and battery. Display size is what tells when you’re reading the long answers an AI hands back. The Duo’s 7.6 inches is a different thing from the 18 Pro’s 6.3. I personally got tired enough of reading generated text that I started sending it to an e-paper device to read instead. Having more area to read on changes the quality of that work all by itself.

What I Haven’t Confirmed in This Article

• I don’t own the hardware. There isn’t a single speed number here
• The amount of memory is not disclosed by Apple. The 12GB / 9GB figures for the 18 series are estimates, and the base iPhone 18 in particular is split between a 9GB and a 12GB prediction
• The 17 series’ 8GB / 12GB figures and the iOS 27 feature split come from developer tooling and various write-ups, not from Apple’s specs pages
• Details on “20B runs on-device" could not be confirmed on the official page. This comes from various write-ups
• Whether MLX, a separate framework, uses the Neural Engine — I could not confirm this
• The specific memory ceiling for a single iOS app is not something I could confirm on Apple’s official pages. The guidelines for shipping a small AI model in an app (2-billion class, 15–30 tokens per second, about 44% drop from heat) all come from various write-ups
• The three-phone comparison comes from Apple’s official specs pages, though the amount of memory is stated for none of them

Closing — Look at the “1.5x," Not the “2x"

What doubled is the number of Neural Engine cores. Two 16-core Neural Engines now sit on the chip for 32 combined, and that is what Apple calls “2x the AI performance." It does not say speed doubles. Core count and the speed you get out are different things.

And the Neural Engine is reached mainly through Apple’s own framework. Running an AI model you chose yourself doesn’t go through it. This “2x" is the same whichever of the three phones you pick.

What matters is the other figure announced at the same time: memory bandwidth at 1.5x. Since write speed comes down to a bandwidth division, this one feeds straight back into actual speed.

And the number I most want to know — the amount of memory — isn’t disclosed. Without that figure fixed, there’s no way to know which AI models will even fit.

All three phones share the chip and the AI circuitry. I found no reason to pick one for speed. What differs is display and battery, and the Duo in particular won’t run anything faster, but the area you have to read the returned text on nearly doubles. When you’re running AI locally, the reading side stays a human job no matter what. Seen as a device that helps with the reading side, it looks a little different.