- that a bunch of users won't jump ship if Apple stagnates for 5 years
- that a product based on a model with Q12026 SoTA performance would be competitive with products using 2031's models.
- that just having access to good (by 2025/2026 standards) models is the big thing that Apple needs in order for Apple Intelligence to finally be useful.
On that last point, I think the OS/app-level features are almost more important than the model itself. If the model can't _do_ anything, it doesn't really matter how intelligent it is. If Apple sits on their laurels for 5 years, would their OS, built-in apps, and 3rd-party apps have all the hooks needed for a useful AI product?This is the most unlikely proposition, given how Apple has managed to be a decade behind and still very profitable.
Besides that, having a Claw-like AI with full access to your phone is surely a recipe for disaster. IMO Apple is being justifiably cautious in staying an spectator, looking busy, and waiting to make a deal with the winner of the chatbot wars.
A good analogy would be streaming. To get good quality, sure, you can store the video file but it is going to take up space. For videos, these are 2-4GB (lets say) and streaming will always be easier and better.
For models, we're looking at 100s of GB worth of model params. There's no way we can make it into, say, 1GB without loss in quality.
So nope, beyond minimal classification and such, on-device isnt happening.
--
EDIT:
> Nobody wants to be sending EVERY request to someone else's cloud server.
We do this already with streaming. You watch YouTube that is hosting videos on the "cloud". For latest MKBHD video, I dont care about having that locally (for the most part). I just wanna watch the video and be done with it.
Same with LLMs. If LLMs are here to stay, most people would wanna use the latest / greatest models.
---
EDIT-EDIT:
If you response is Apple will figure it out somehow. Nope, Apple is sitting out the AI race. So it has no technology. It has nothing. It has access to whatever open source is available or something they can license from rest. So nope, Apple isnt pushing the limits. They are watching the world move beyond them.
Indeed.
But they said 5 years. That's certainly plausible for high-end mobile devices in Jan 2031.
I have high uncertainty on if distillation will get Opus 4.6-level performance into that RAM envelope, but something interesting on device even if not that specifically, is certainly within the realm of plausibility.
Not convinced Apple gets any bonus points in this scenario, though.
And anyway, you already see models like Qwen 3.5 9B and 4B beating 30B and 80B parameter models, which can already run on phones today, especially with quantization.
Benchmarks: https://huggingface.co/Qwen/Qwen3.5-4B
This is a paradox right? Handset makers want less handset storage so they can get users to buy more of their proprietary cloud storage while at the same time wanting them to use their AI more frequently on their handsets.
It will be interesting which direction they decide to go. Finding a phone in the last few years with more than 256gb storage is not only expensive AF, its become more of a rarity than commonplace. Backtracking on this model in order to simply get AI models on board would be a huge paradigm shift.
Useful LLM usage involves pushing a lot of private data into them. There's a pretty big difference sending up some metadata about your viewing of an MKBHD video, and asking an LLM to read a text message talking about your STD test results to decide whether it merits a priority notification. A lot of people will not be comfortable with sending the latter off to The Cloud.
1/ No, you don't get Opus 4.6 level on devices with 12Gb of RAM, 7B quantised models just don't get that good. Still quite good mind you, and I believe that the biggest advance to come from mobile AI would be apps providing tools and the device providing a discovery service (see Android's AppFunctions, if it was ever documented well): output quality doesn't matter on device, really efficient and good tool calling is a game changer.
2/ Opus 4.6 is now Opus 4.6+5years and has new capabilities that make people want to keep sending everything to someone else's cloud server instead of burning their battery life
And that is exactly why it won't happen (like that).
This is not a huge disadvantage in my opinion. Let the rest of big tech fight each other to death over cloud, while controlling a very profitable differentiated offering (devices+services). Apple keeps the M series HW out of data centers, even though it presents some very attractive performance/w and per-core numbers.
What I really want is my phone to transcribe all of my phone calls to a Notes document. Since it isn't recording an audio conversation, I don't think the consent laws come into play.
>What I really want is my phone to transcribe all of my phone calls to a Notes document
This has been doable for decades. Why haven't you done it? My Pixel phones did this with voicemail before LLMs.
Windows Vista shipped with full featured dictation functionality, and it works better than you would expect, all local, all using classical algorithms, all evaluated cheaply. If it wasn't accurate enough, Dragon speech to text tools were gold standard for most of modern computing history, and greatly surpassed the accuracy of that built in system.
BTW, you can, on any Windows machine right now, access that built in voice recognition, and with a "Constrained vocabulary", say if you only want a few specific voice commands, it gets near perfect accuracy constantly. You have to search for old documentation now because Microsoft wants to hide that you don't need an internet connection or an Azure account and monthly bill to ship accurate voice recognition with your app. It's trivial to use, from both C++ and C#, and anything else that allows you to invoke native code, and the workflow is easy enough to understand. I built an app to utilize it instead of buying one of those $10 "Voice control your game" apps to add voice control to ARMA, and it was easier to implement the voice recognition than it was to copy and paste native code invocations for the Win32 api to inject keystrokes. I don't even write C# code in general.
https://learn.microsoft.com/en-us/previous-versions/windows/...
There's tons of documentation about "Grammar" and configuration but the default configuration IIRC is to just turn speech input into text, and do so with at least 85% accuracy, even without the user actually training the recognizer to their voice. If you build context specific grammars or a hierarchical grammar to support a real UX that isn't just hoping some code knows how to interpret raw speech you will get dramatically better recognition performance.
This is IMO a frequent pattern. Time and time again the people who keep saying "I want LLMs to do X" don't seem to be aware that "X" was a robust and mature area of research decades ago! They don't seem to be aware that you could already do X and even buy ready to go software for that purpose! Often enough the LLM version is an outright regression in functionality, as things that were doable with a single microchip in 1960 now require an internet connection.
>Since it isn't recording an audio conversation,
So to be clear, you want this functionality explicitly to bypass law? Federally and in 39ish states, you only need your own consent anyway.
It would at least be nice if they could do some basic attempts to detect scam texts, because I frequently see AI message summaries about supposed $2000 purchases made on my Amazon account, usually worded more eloquently than the original message.
The Siri+LLM features of Apple Intelligence aren’t launched yet, and the other features like notification summaries run on-device.
Just a total failure of execution.
Siri being better at free form requests for actions and doing internet/knowledge searches is about all I can think of. But also, I use Kagi for that, and unless Siri has a pluggable backend for search I'm not sure being forced to use only Apple's search, if it ever exists, is a great design.
If it even works, it fails with "something went wrong" for me 3 out of 5 times
https://www.macrumors.com/2026/01/30/apple-explains-how-gemi...
It somehow looks worse than most scammy image generation apps you see on half-page search ads on the App Store. I have no idea how Apple willingly released it like that.
It was updated on my iPhone to a bland, forgettable abstract icon that’s still fairly mediocre but no longer an ongoing embarrassment for their corporate brand standards.