[HN Gopher] Apple Intelligence Foundation Language Models Tech R...
___________________________________________________________________
Apple Intelligence Foundation Language Models Tech Report 2025
Author : 2bit
Score : 130 points
Date : 2025-07-17 18:09 UTC (4 hours ago)
(HTM) web link (machinelearning.apple.com)
(TXT) w3m dump (machinelearning.apple.com)
| leot wrote:
| The dozens of "contributors" being presented in random order is,
| one would suppose, an anti-poaching tactic?
| zamadatix wrote:
| It's hard to know what it isn't for certain but there are many
| other reasons papers list contributors in a flat structure (be
| it random or alphabetical order). Particularly with large
| numbers of collaborators.
| JKCalhoun wrote:
| "References" section sort of narrows the field anyway.
| rafram wrote:
| Not very hard to look people up on LinkedIn and figure out who
| the core researchers are. I think this is just a very surface-
| level overview paper that encompasses a bunch of different
| research projects conducted by different teams, and it would be
| difficult to order the contributors in any meaningful way.
| browningstreet wrote:
| As someone whose last name is near the end of the alphabet,
| that's not the first presumption I had seeing that page.
| ml-anon wrote:
| Well meta already got Ruoming so he can obviously give them a
| ranked list of who to grab.
|
| Most of his team are former Google brain so GDM knows who is
| good.
| JKCalhoun wrote:
| I wonder if we'll see these models running on the phone (aiPhone)
| hardware in the future.
| kingnothing wrote:
| > The new Foundation Models framework gives access to
| developers to start creating their own reliable, production-
| quality generative AI features with the approximately 3B
| parameter on-device language model. The ~3B language foundation
| model at the core of Apple Intelligence excels at a diverse
| range of text tasks like summarization, entity extraction, text
| understanding, refinement, short dialog, generating creative
| content, and more. While we have specialized our on-device
| model for these tasks, it is not designed to be a chatbot for
| general world knowledge. We encourage app developers to use
| this framework to design helpful features tailored to their
| apps
| Zee2 wrote:
| > a ~3B-parameter on-device model
| ThomasBb wrote:
| There are even already some local AFM to Open AI API bridge
| project on GitHub - that lets you point basically any Open AI
| compatible client at the local models. Super nice for basic
| summarisation and completions.
| JKCalhoun wrote:
| I was worried "device" was a Mac mini, not an iPhone. (I
| already have been running models on my MacBook Pro.)
| floam wrote:
| It does. You can use it directly on iOS 26 beta - without
| writing a line of code I can toy with the on-device model
| through Shortcuts on my 16 Pro. It's not meant to be a general
| purpose chatbot... but it can work as a general purpose chatbot
| in airplane mode which is a novel experience.
|
| https://share.icloud.com/photos/018AYAPEm06ALXciiJAsLGyuA
|
| https://share.icloud.com/photos/0f9IzuYQwmhLIcUIhIuDiudFw
|
| The above took like 3 seconds to generate. That little box that
| says On-device can be flipped between On-device, Private Cloud
| Compute, and ChatGPT.
|
| Their LLM uses the ANE sipping battery and leaves the GPU
| available.
| bigyabai wrote:
| It would be interesting to see the tok/s comparison between
| the ANE and GPU for inference. I bet these small models are a
| lot friendlier than the 7B/12B models that _technically_ fit
| on a phone but won 't accelerate well without a GPU.
| gleenn wrote:
| I thought the big difference between the GPU and ANE was
| that you couldn't use the ANE to train. Does the GPU
| actually perform faster during inference as well? Is that
| because the ANE are designed more for efficiency or is
| there another bigger reason?
| wmf wrote:
| GPUs are usually faster for inference simply because they
| have more ALUs/FPUs but they are also less efficient.
| JKCalhoun wrote:
| Wild to see what improvements might come if there is
| additional hardware support in future Apple Silicon chips.
| ivape wrote:
| What's the cost of pointing it to Private Cloud Compute? It
| can't be free, can it?
| alwillis wrote:
| As someone mentioned, this model is available in the beta
| version of iOS 26; it's also part of macOS 26, iPadOS 26 and
| visionOS 26. Anyone with a free developer account can install
| the developer betas; the public beta is expected next week.
|
| There's a WWDC video "Meet the Foundation Models Framework"
| [1].
|
| [1]: https://developer.apple.com/videos/play/wwdc2025/286
| jonplackett wrote:
| Every time I see a paper from Apple I just feel like, OK so why
| isn't my iPhone actually doing any of this yet?
|
| Why give this to developers if you haven't been able to get Siri
| to use it yet? Does it not work or something? I guess we'll find
| out when devs start trying to make stuff
| bayindirh wrote:
| > why isn't my iPhone actually doing any of this yet?
|
| Probably Apple is trying to distill the models so they can run
| on your phone _locally_. Remember, most, if not all, of Siri is
| running on your device. There 's no round trip whatsoever for
| voice processing.
|
| Also, for larger models, there will be throwaway VMs per
| request, so building that infra takes time.
| jonplackett wrote:
| It says there's 2 models - one local. It's already released
| to app developers to use locally I think (it was in the
| keynote for WWDC).
| geoffpado wrote:
| The model now available to developers (in beta, not in
| released versions of iOS) is the same model that powers
| stuff like the much-maligned notification summaries from
| iOS 18. So your phone _does_ have features that are powered
| by this stuff... you may just not be particularly
| overwhelmed by those features.
| jonplackett wrote:
| That's kinda my point though - is this only capable of
| things like this? If it ia capable of more, why isn't
| there something more yet, it's been a long time
| waiting...
| imoverclocked wrote:
| > why isn't my iPhone actually doing any of this yet?
|
| What exactly are you referring to? Models do run on iPhone and
| there are features that take advantage of it, today.
| jonplackett wrote:
| None of those features are in any way interesting though.
| Image playground is a joke, Siri is a joke, that generative
| emoji thing is a joke.
|
| The AI stuff with photography sure, but that's more like
| machine learning.
|
| The photo touch up thing is... useable? Sometimes?
|
| What is it you've been so impressed with?
| perfmode wrote:
| > We believe in training our models using diverse and high-
| quality data. This includes data that we've licensed from
| publishers, curated from publicly available or open- sourced
| datasets, and publicly available information crawled by our web-
| crawler, Applebot.
|
| > We do not use our users' private personal data or user
| interactions when training our foundation models. Additionally,
| we take steps to apply filters to remove certain categories of
| personally identifiable information and to exclude profanity and
| unsafe material.
|
| > Further, we continue to follow best practices for ethical web
| crawling, including following widely-adopted robots.txt protocols
| to allow web publishers to opt out of their content being used to
| train Apple's generative foundation models. Web publishers have
| fine-grained controls over which pages Applebot can see and how
| they are used while still appearing in search results within Siri
| and Spotlight.
|
| Respect.
| bitpush wrote:
| When Apple inevitably partners with OpenAI or Anthropic, which
| by their definition isnt doing "ethical crawling", I wonder how
| I should be reading that.
| jhickok wrote:
| They already partnered with OpenAI, right?
| DSingularity wrote:
| To use their APIs at a discount, so what?
| JimDabell wrote:
| Apple aren't paying OpenAI anything:
|
| https://www.bloomberg.com/news/articles/2024-06-12/apple-
| to-...
| sharkjacobs wrote:
| That's a big discount
| badwolf wrote:
| That's quite a discount! ;)
| fridder wrote:
| Same way as the other parts of their supply chain I suppose.
| wmf wrote:
| In theory Apple could provide their training data to be used
| by OpenAI/Anthropic.
| bitpush wrote:
| It isn't "apple proprietary" data to give it to OpenAI.
|
| Also the bigger problem is, you can't train a good model
| with smaller data. The model would be subpar.
| simonw wrote:
| One problem with Apple's approach here is that they were
| scraping the web for training data long before they published
| the details of their activities and told people how to exclude
| them using robots.txt
| dijit wrote:
| Uncharitable.
|
| Robots.txt is _already_ the understood mechanism for getting
| _robots_ to avoid scraping a website.
| simonw wrote:
| People often use specific user agents in there, which is
| hard if you don't know what the user agents are in advance!
| wat10000 wrote:
| If you're using a specific user agent, then you're saying
| "I want this specific user agent to follow this rule, and
| not any others." Don't be surprised when a new bot does
| what you say! If you don't want any bots reading
| something, use a wildcard.
| JacobJack wrote:
| I'd really like to be able to use this 3B model on my little 4GB
| GPU card! It looks very capable for a reasonable weight. Maybe
| one day on HhuggingFace
| frankfrank13 wrote:
| AFAICT this is the first commercial model _trying_ to be marketed
| as responsibly-sourced. Love it, but it also seems like the noise
| around this issue has died down. Is this for legal cover? Or more
| apple-privacy marketing
| Daedren wrote:
| Stockholders are suing them over Apple Intelligence. Definitely
| legal cover.
| msgodel wrote:
| Do people have an issue with the smollm datasets? I guess it
| isn't really commercial.
| jiehong wrote:
| Looks nice. I just wish they'd improve the models behind
| dictation on both iPhone and Mac to have better accuracy and on
| the fly multiple language transcription.
| poszlem wrote:
| In the meantime, when I ask Siri to set a timer for 15 minutes,
| about 10-15% of the time it just says, "Here's what I found about
| setting a timer for 15 minutes," instead of actually setting the
| timer"
| mittermayr wrote:
| All I can say is, I asked Siri today (verbatim): What is 75
| degrees fahrenheit in celsius, and what is 85 degrees in
| fahrenheit -- and it offered a web search about fahrenheit. The
| "and" completely disabled its most basic ability to do metric
| conversions.
|
| So, it's nice to see Apple is doing research and talking about
| it, but we're out here waiting, still waiting, for anything
| useful to make of it all on our thousand-dollar devices that
| literally connect us to the world and contain our entire life
| data. It's what I would've expected from one of the most valuable
| companies in the world.
| basisword wrote:
| >> What is 75 degrees fahrenheit in celsius, and what is 85
| degrees in fahrenheit
|
| Probably wouldn't have made a difference but the second half of
| that statement isn't exactly clear. 85 degrees what?
|
| I also think when you're chaining these two separate
| calculations together you get a problem when it comes to
| displaying the results.
| bronco21016 wrote:
| Meanwhile users have been conditioned to expect a system that
| understand the multiple queries and answers them
| appropriately.
| JKCalhoun wrote:
| True. But for most of us, only in the past year. I have a
| few friends/relatives who have still never conversed with
| an LLM.
| vosper wrote:
| That exact phrase "What is 75 degrees fahrenheit in celsius,
| and what is 85 degrees in fahrenheit" given to ChatGPT
| produces the correct result (it infers that the second
| degrees must be Celsius) and ChatGPT gives me a nicely laid
| out formula for the math of the conversion.
|
| So yeah, Apple is way behind on this stuff.
| seydor wrote:
| the fact is that gemini responds with this: 75 degrees
| Fahrenheit is 23.89 degrees Celsius, and 85 degrees Celsius
| is 185.00 degrees Fahrenheit.
| losvedir wrote:
| > _What is 75 degrees fahrenheit in celsius, and what is 85
| degrees in fahrenheit_
|
| Err, what? As a native English speaker human that's a pretty
| confusing question to me, too!
| sneilan1 wrote:
| I feel like this is the most exciting news today about AI on hn.
| I really hope apple shows that small models can be just as
| capable as the bigger ones. Maybe they have the people on
| perplexity working on these small models.
| bitpush wrote:
| The more I think about Apple, the more I realize that Apple is so
| far behind. While other companies are pushing the envelope
| (OpenAI, Anthropic, Google ..) Apple's ambitions seem much much
| smaller.
|
| And this is after they made very big claims with Apple
| Intelligence last year, when they had everyone fooled.
|
| This is like watching a train-wreck in slow motion.
| outworlder wrote:
| Only if you think they _must_ compete with large models on the
| internet.
| Uehreka wrote:
| I wouldn't go as far as GP, but yes, absolutely, they must
| compete with large models on the internet. Customers are now
| used to being able to ask a computer a question and get
| something better than "I just ran a web search for what you
| said, here are the uncurated, unsummarized results".
|
| Yes, this is in fact what people want. Apple is the biggest
| company in the world (don't quibble this y'all, you know what
| I mean) and should be able to deliver this experience. And
| sure, if they could do it on device that would be aces, but
| that's not an item on the menu, and customers seem fine with
| web-based things like ChatGPT for now. To act like Apple is
| doing anything other than fumbling right now is cope.
| GeekyBear wrote:
| > I wouldn't go as far as GP, but yes, absolutely, they
| must compete with large models on the internet
|
| The people running large models want to charge a monthly
| fee for that.
|
| I'm fine with having a free model that runs on device
| without slurping up my data.
| specialist wrote:
| I'm fine with Apple chilling on the sidelines for a bit.
| JKCalhoun wrote:
| I see it as the opposite. Apple is absolutely positioned to own
| "chat". I am not worried they'll soon sort things out -- and
| eventually we'll have an LLM integrated into the iPhone; call
| it Siri or otherwise.
|
| With my history encrypted in the cloud, and the trust that
| Apple has built around privacy ... I think they're going to
| come out alright.
| martinald wrote:
| But they have de facto admitted failure of most of the
| strategy if the rumours are true that they are switching much
| harder to OpenAI/Anthropic for upcoming LLM products.
|
| This is the first time in 10+ years I've seen Apple so far on
| the back foot. They usually launch category defining products
| that are so far ahead of the competition, even by the time
| they work through the 'drawbacks' in the first versions of
| them they are still far ahead. OS X, the iPhone and the iPad
| were all like that. They are still way ahead of the
| competition on Apple Silicon as well.
|
| I am not very confident on their on device strategy at least
| in the short to medium term. Nearly all their devices do not
| have enough RAM and even if they did SLMs are very far behind
| what users "know" as AI - even the free ChatGPT plan is leap
| years ahead of the best 3B param on device model. Maybe there
| will be huge efficiency gains.
|
| Private cloud is used AFIAK for virtually 0 use cases so far.
| Perhaps it will be more interesting longer term but not very
| useful at the moment given the lack of a suitable (ie: non
| Chinese), large (>500b param) model. They would also struggle
| to scale it if they roll it out to billions of iOS devices
| especially if they put features that use a lot of tokens.
|
| Then they've got OpenAI/Gemini/Anthropic via API. But this
| completely goes against all their private cloud messaging and
| gives those providers enormous potential control over Apple,
| which is not a position Apple usually finds itself in. It
| will also be extremely expensive to pay someone per token for
| OS level features for billions of iOS/Mac devices and unless
| they can recoup this via some sort of subscription will hit
| services margins badly.
|
| To me its clear the future of "OS" is going to involve a lot
| of agentic tool calling. These require good models, with
| large context windows and a lot of tokens - this will
| definitely not work on device. Indeed this is exactly what
| the Siri vapourware demo was.
|
| I'm sure they can potentially get to a great UX (though these
| missteps are making me question this). But having such a core
| feature outsourced does not leave them in a good position.
| visarga wrote:
| The paper was a very nice read, and they did many creative
| things. It's a pity this model won't be directly accessible,
| only integrated in some apps.
| alwillis wrote:
| > It's a pity this model won't be directly accessible, only
| integrated in some apps.
|
| It's already accessible using Shortcuts, even to non-
| developers "iOS 26 Shortcuts + Apple Intelligence is POWERFUL
| " (Youtube) [1].
|
| [1]: https://youtu.be/Msde-lZwOxg?si=KJqTgtWjpdNDxneh
| halJordan wrote:
| Apple's ambitions are actually bigger than openai or
| anthropopic. Only Google's ambition (surprise surprise) is
| similar. Apple fundamentally wants the llm to be a tool. It
| doesn't want the llm to be the product.
| chevman wrote:
| Siri is literally a joke!
|
| My son (he's 11 years old now and fairly skilled with all the
| main AI tools, eg chatgpt, gemini, etc) and I retry her every
| month or so, and this past time we just laughed. Can't handle
| basic questions - hears the question wrong, starts, stops, takes
| us to some random ass webpage, etc, etc.
|
| "She's so jacked up!" he said.
|
| Apple needs to get this under control and figured out, stat!
| EGreg wrote:
| Here is my question...
|
| This is the first time that millions of people will actually
| download and run a model on their own devices.
|
| The question is... will Apple be constantly tweaking these
| models, or only during OS upgrades?
|
| I for one really like local software. Call me old-fashioned, but
| I enjoy when a company doesn't switch up software anytime on the
| server, or phone the results home all the time in order to
| extract more profits from their collective users.
| wmf wrote:
| The model is gigabytes so I doubt they will push updates
| frequently.
| ApolloVonZ wrote:
| Despite all the "Apple is evil" or "Apple is behind" (because
| they don't do evil). Well, what they made with the Foundation
| Model is great. The fact that they build a system within the
| Swift language that allows you to specify structured data models
| (structs) to be used like any other model in a modern programming
| language, and you actually get back generated data in that format
| is great. Unlike a lot of other AIs where you might get back a
| well formatted JSON after a carefully crafted request, but still
| you never can't be sure and need to implement a bunch of
| safeguards. Obviously it's still the beginning and other tools
| might do something similar. But as an iOS developer that makes
| the usage of AI so much simpler. Especially with the bridge to
| external AIs that still allows you to map back to the type safe
| structured Swift models. I try not to be a hater, every progress,
| even slow or underwhelming at first might lead to improvements
| everywhere else.
| 0x457 wrote:
| Guided generation is called "Structured Output" by other
| providers?
|
| Well partially generated content streaming thing is great and I
| haven't seen it anywhere else.
___________________________________________________________________
(page generated 2025-07-17 23:00 UTC)