Apple Can Create Smaller On-Device AI Models From Google's Gemini

[Wx_Man]

The mainstream models were created by ingesting the world's data. It didn't make sense for Apple to go and do that themselves when it's already been done, especially by Google which has excelled at that.

Apple is all about bringing a technology to the masses in a friendly, efficient way. The long game.

They didn't invent the graphical interface. They perfected it.

They didn't invent the mouse. They perfected it.

They didn't invent the portable computer. They perfected it.

They didn't invent the digital music player. They perfected it.

They didn't invent the smartphone. They perfected it.

... and now we have the AI experience, ready to be perfected.

“Well….like…that’s just your opinion, man.”

The Dude

 

[Wx_Man]

Ok, but how are they going to fit those LLMs + the OS into 8GB of RAM?

Elfin Magic.

That is if Apple can negotiate the terms that Keebler desires for a buyout.

 

[Wx_Man]

So is Apple paying Google $5 billion for access and distillation?

I’ve read $ 1 Billion per year. Either way it’s simply a deduction from the $ 20-25 Billion Google pays Apple every year to be the default search engine for Safari.

 

[Wx_Man]

Absolutely
View attachment 2616762

Yeah, the perfected mouse. The one that becomes useless once you need to charge it AND use it at the same time.

 

[jillownsamac]

ewww...
Seeing how Google's view on privacy (looking at the incognito mode incident), I would not trust them one bit.
They would have a back door to the data. 🙁

What are the privacy implications of this?

Also what the heck does this even mean (and mean re privacy) : "Google gave Apple "complete access" to the Gemini model in its own data centers, and Apple can use the access for distillation, or creating smaller models for specific tasks." (quote this macrumors article)

 

[HazeAndHahmahneez]

Absolutely
View attachment 2616762

Nice try but that’s actually a great mouse, it’s just many ppl can’t use it lol

…and surely you know he’s prob referencing the og Mac mouse from what the 80s? 90s?

Do not conflate Gemini chat with the Gemini model. Same for Chat GPT and GPT 5.xx

The chat bits are a layer on top of the model. This is the way that Siri will work. Copilot is similar.

You can use the model 100% obfuscated from the company that created it. This is how Apple will insure user privacy. Nothing gets back to Google.

Let’s hope and pray nothing ever does but even so, I got Siri and Apple Intelligence turned ALL THE WAY OFF, I rec. everyone else does as well for the foreseeable fyootch.!

 

[eifelbube]

…
They didn't invent the graphical interface. They perfected it.
They didn't invent the mouse. They perfected it.
They didn't invent the portable computer. They perfected it.
…

They did invent the OS 26 graphical user interface. They blew it.

... and now we have the AI experience, ready to be perfected.

> > Siri, directions to Whole Foods
One option I found is Dunkin’ Donuts on South Milton Road, about 5 miles to your north, and it gets 68% positive ratings. Does that one sound good?
> > No
The second is Fat Bagels on Historic Route 66, about 7 miles to your east, and it gets 2.9 stars on Yelp. Do you want that one?

P.S.
You forgot:
They didn’t invent the tablet computer. They …

 

[McScooby]

Claude is exceptional, too. I do hope that "a new Apple" has the wisdom to be more open, voluntarily, rather than being forced to be.

Is Claude "invisible" in its usage yet? Siri will be. That's Apple's end game. Using technology without knowing you're using technology.

I'm sure it would be if I was mad enough to give it full access and permissions!😂

 

[H2SO4]

The mainstream models were created by ingesting the world's data. It didn't make sense for Apple to go and do that themselves when it's already been done, especially by Google which has excelled at that.

Apple is all about bringing a technology to the masses in a friendly, efficient way. The long game.

They didn't invent the graphical interface. They perfected it.

They didn't invent the mouse. They perfected it.

They didn't invent the portable computer. They perfected it.

They didn't invent the digital music player. They perfected it.

They didn't invent the smartphone. They perfected it.

... and now we have the AI experience, ready to be perfected.

I'm not sure you know what 'perfected' means.

 

[eifelbube]

Absolutely
View attachment 2616762

They didn’t invent wireless mouse charging. They perverted it.

 

[UltimaKilo]

I think it is pretty clear what this has to do with this news. If you think Gemini is the best LLM on the market, I don't think you have very much experience with its competitors.

What is “clear” is that you came in here to troll.

Secondly, it’s the best LLM by almost every metric. Only Grok beats it out with less hallucinations and more up-to-date news results. OpenAI are expecting a major update in May to close the gap with Gemini and xAI expected to respond by June. No major news on Anthropic, yet.

• ARC-AGI-2 (Logic): Gemini 3.1 Pro (77.1%) > Opus 4.6 (68.8%) > GPT-5.2 (52.9%).
• GPQA Diamond (Science): Gemini 3.1 Pro (94.3%) > GPT-5.2 (92.4%) > Opus 4.6 (91.3%).
• MMLU (Understanding): Gemini 3.1 Pro (92.6%) > Opus 4.6 (91.1%) > GPT-5.2 (89.6%).
• Humanity's Last Exam: Gemini 3.1 Pro (44.4%) > Opus 4.6 (40.0%) > GPT-5.2 (34.5%).

Gemini 3.1 Pro vs Claude Opus 4.6 vs GPT-5.2: Best AI… | NxCode

Gemini 3.1 Pro vs Claude Opus 4.6 vs GPT-5.2 compared across benchmarks, pricing, coding, and reasoning. Find the best AI model for your needs.

www.nxcode.io

Google releases Gemini 3.1 Pro: Benchmarks, how to try it

Google says that its most advanced thinking model yet outperforms Claude and ChatGPT on Humanity's Last Exam and other key benchmarks.

mashable.com

 

[User 6502]

First of all, what does that have to do with this news? Secondly, your "testing" must have some serious flaws because currently, Gemini is the best LLM on the market. That, of course, can change very soon.

What tests? Claude seems a lot better to me and it beats Gemini in many metrics.

 

[Joe Bogger]

They did invent the OS 26 graphical user interface. They blew it.


> > Siri, directions to Whole Foods
One option I found is Dunkin’ Donuts on South Milton Road, about 5 miles to your north, and it gets 68% positive ratings. Does that one sound good?
> > No
The second is Fat Bagels on Historic Route 66, about 7 miles to your east, and it gets 2.9 stars on Yelp. Do you want that one?

P.S.
You forgot:
They didn’t invent the tablet computer. They …

You are absolutely right. Siri is missing the point. It should had also added: this is the option to the diabetic treatment center where you go afterwards. Does that one sound good?

 

[Joe Bogger]

They didn’t invent wireless mouse charging. They perverted it.
View attachment 2616932

this is called Dead Mouse

 

[Joe Bogger]

By the time Apple will arrive in the chat bot area everyone else will be off in the agent area.

 

[leonremi]

getting "emotional support" from a chatbot is about the worst thing anyone could do with it.

 

[sevoneone]

Wouldn't this just be gemma3:1b ?

This is my question too. Gemma 3 is based on Gemini 2.0 though, so I am guessing what Apple is working with is the next generation based on Gemini 3.x. I assume Apple wants on device Siri to be multimodal. Gemma 3 270m and 1b are text only, the ‘efficient’ 2b version needs at least 2GB of memory which I think would be a rough fit on an 8GB device like the Neo without impacting performance.

Apple needs needs a really basic win with the first iteration of on-device Siri: Fast, accurate, actually useful. Multimodal: text, voice, image and video input. Text and voice output. No more “playgrounds” or experiments.

As the memory available increases i expect these models to be more and more useful.

Memory available increasing does’t seem very likely given prices and scarcity of RAM currently.

 

[jeffbax]

I almost don’t care what model they license at this point, I just hope Apple is able to iterate faster now and deliver something useful.

They are wise on their capex spend, but embarrassed given the speed of change in capability of these tools and now agents in the past few years that have made my head spin.

They need a new release strategy, yearly monolithic software doesn’t work with delivering quality capability.

 

[leonremi]

I almost don’t care what model they license at this point, I just hope Apple is able to iterate faster now and deliver something useful.

They are wise on their capex spend, but embarrassed given the speed of change in capability of these tools and now agents in the past few years that have made my head spin.

They need a new release strategy, yearly monolithic software doesn’t work with delivering quality capability.

I'm honestly confused. Where is the true leap of capability in real world usage?

 

[michaeljk]

It's not possible for Google to have a back door with this. It's just a model that stands alone. It's like going to a farmer's market and buying an apple. You can do whatever you want with the apple once you buy it. The farmer who grew it on a tree doesn't have access to anything you do with the apple.

I am not sure I trust this. I do hope Apple provides some clarity around this. I seem to remember recently seeing a headline that Apple was exploring using Google's servers for some of the Gemini functionality it was porting over. One thing I liked about Apple's initial plans for Apple Intelligence was the hierarchy of privacy protection with three layers. First, on device AI (yours only). Second, Apple "cloud" (or whatever they called it) servers to do larger queries and tasks, which would only use Apple's servers that had robust privacy protection. Third, standard ChatGPT (you decide whether you want to trust OpenAI to keep your work with it private because it leaves your device and it leaves Apple's servers). I hope they do something similar with Gemini, or better yet, they don't even include the third option in whatever they do and they allow you to decide who you trust (right now for keeping my work with AI private, I do trust Apple, and I mistrust the least Anthropic, and then Perplexity, and I don't trust at all OpenAI, Microsoft, Google, Meta, Grok and other foreign models). Mostly, who I trust has to do with their economic model. If it is ad based, partially or fully, no trust whatsoever--they survive by tracking, keeping, selling and mining your data. If it is suscription based (for now), then my trust is middling (Anthropic and Perplexity). If the company is hardware and subscription based and privacy is a main selling point, and they have fought governments and possible revenue partners to their own detriment to keep user data private (Apple and only Apple as far as I know), then I trust them.

I hope we get to a point where we have devices that are powerful enough so they can, like @neuropsychguy says, download and contain a complete model, on device, that can then go out into the internet to do its own work, including searches in private, without having to worry about who is following or keeping memories of the searches.

 

[Sputman99]

Does that make it easier to turn all the AI off

 

[PBG4 Dude]

I am not sure I trust this. I do hope Apple provides some clarity around this. I seem to remember recently seeing a headline that Apple was exploring using Google's servers for some of the Gemini functionality it was porting over. One thing I liked about Apple's initial plans for Apple Intelligence was the hierarchy of privacy protection with three layers. First, on device AI (yours only). Second, Apple "cloud" (or whatever they called it) servers to do larger queries and tasks, which would only use Apple's servers that had robust privacy protection. Third, standard ChatGPT (you decide whether you want to trust OpenAI to keep your work with it private because it leaves your device and it leaves Apple's servers). I hope they do something similar with Gemini, or better yet, they don't even include the third option in whatever they do and they allow you to decide who you trust (right now for keeping my work with AI private, I do trust Apple, and I mistrust the least Anthropic, and then Perplexity, and I don't trust at all OpenAI, Microsoft, Google, Meta, Grok and other foreign models). Mostly, who I trust has to do with their economic model. If it is ad based, partially or fully, no trust whatsoever--they survive by tracking, keeping, selling and mining your data. If it is suscription based (for now), then my trust is middling (Anthropic and Perplexity). If it is hardware and subscription based and privacy is a main selling point, and have fought governments and possible revenue partners to your own detriment to keep user data private (Apple and only Apple as far as I know), then I trust you.

I hope we get to a point where we have devices that are powerful enough so they can, like @neuropsychguy says, download and contain a complete model, on device, that can then go out into the internet to do its own work, including searches in private, without having to worry about who is following or keeping memories of the searches.

It’s coming. Alex Ziskind did a video of Tiny ai’s Pocket AI device connected to a MacBook Neo.

 

[iMac The Knife]

What is “clear” is that you came in here to troll.

Secondly, it’s the best LLM by almost every metric. Only Grok beats it out with less hallucinations and more up-to-date news results. OpenAI are expecting a major update in May to close the gap with Gemini and xAI expected to respond by June. No major news on Anthropic, yet.

• ARC-AGI-2 (Logic): Gemini 3.1 Pro (77.1%) > Opus 4.6 (68.8%) > GPT-5.2 (52.9%).
• GPQA Diamond (Science): Gemini 3.1 Pro (94.3%) > GPT-5.2 (92.4%) > Opus 4.6 (91.3%).
• MMLU (Understanding): Gemini 3.1 Pro (92.6%) > Opus 4.6 (91.1%) > GPT-5.2 (89.6%).
• Humanity's Last Exam: Gemini 3.1 Pro (44.4%) > Opus 4.6 (40.0%) > GPT-5.2 (34.5%).

Gemini 3.1 Pro vs Claude Opus 4.6 vs GPT-5.2: Best AI… | NxCode

Gemini 3.1 Pro vs Claude Opus 4.6 vs GPT-5.2 compared across benchmarks, pricing, coding, and reasoning. Find the best AI model for your needs.

www.nxcode.io

Google releases Gemini 3.1 Pro: Benchmarks, how to try it

Google says that its most advanced thinking model yet outperforms Claude and ChatGPT on Humanity's Last Exam and other key benchmarks.

mashable.com

First of all, I never troll. Secondly, ChatGPT is widely considered the best LLM on the market today. I can post hundreds of links stating this, but you can do your own research. When it comes to anything recent, Gemini's data is trash by comparison to ChatGPT, Claude, and Grok. Gemini's image generation and manipulation is also garbage compared to ChatGPT and Grok. Gemini didn't even know that Borderland's 4 had been released back in September 2025, until a few weeks ago and depending on the question you ask it will still say the game hasn't been released. With Gemini, I have to repeat the same question over and over again, and only after berating it a few times it will finally come back with a halfway decent answer. This is a drop in the bucket in terms of the failures I've seen from Gemini.

 

[iMac The Knife]

Which would you say is the best?

ChatGPT has been the best I've used. And I hate to say that because I despise Sammy boy.

 

[MrSnirf]

What is “clear” is that you came in here to troll.

Secondly, it’s the best LLM by almost every metric. Only Grok beats it out with less hallucinations and more up-to-date news results. OpenAI are expecting a major update in May to close the gap with Gemini and xAI expected to respond by June. No major news on Anthropic, yet.

• ARC-AGI-2 (Logic): Gemini 3.1 Pro (77.1%) > Opus 4.6 (68.8%) > GPT-5.2 (52.9%).
• GPQA Diamond (Science): Gemini 3.1 Pro (94.3%) > GPT-5.2 (92.4%) > Opus 4.6 (91.3%).
• MMLU (Understanding): Gemini 3.1 Pro (92.6%) > Opus 4.6 (91.1%) > GPT-5.2 (89.6%).
• Humanity's Last Exam: Gemini 3.1 Pro (44.4%) > Opus 4.6 (40.0%) > GPT-5.2 (34.5%).

Gemini 3.1 Pro vs Claude Opus 4.6 vs GPT-5.2: Best AI… | NxCode

Gemini 3.1 Pro vs Claude Opus 4.6 vs GPT-5.2 compared across benchmarks, pricing, coding, and reasoning. Find the best AI model for your needs.

www.nxcode.io

Google releases Gemini 3.1 Pro: Benchmarks, how to try it

Google says that its most advanced thinking model yet outperforms Claude and ChatGPT on Humanity's Last Exam and other key benchmarks.

mashable.com

Benchmarks are nice but actual performance is better. Claude Opus 4.6 > GPT 5.4 > Gemini 3.0 Pro. Gemini 3.1 is worse than 3.0. I've used all of them on my projects and its crazy how good Claude 4.6 is compared to both GPT and Gemini.