Are they doubling down on local LLMs then?
I still think Apple has a huge opportunity in privacy first LLMs but so far I'm not seeing much execution. Wondering if that will change with the overhaul of Siri this spring.
I don't mind it, I open Apple stock. But I'm def not buying into their rebranding of integrated GPU under the guise of Unified Memory.
Aren't the OpenClaw enjoyers buying Mac Minis because it's the cheapest thing which runs macOS, the only platform which can programmatically interface with iMessage and other Apple ecosystem stuff? It has nothing to do with the hardware really.
Still, buying a brand new Mac Mini for that purpose seems kind of pointless when a used M1 model would achieve the same thing.
Yeah because Mac upgrade prices were already sky high, long before the component shortage. 32GB of DDR5-6000 for a PC rocketed from $100 to $500, while the cost of adding 16GB to a Mac was and still is $400.
But if the contract was for a specific amount of RAM and then people start coming to Apple more for high RAM machines, they're going to exhaust their contract sooner than usual and run out of cheap memory to buy. Then they have to decide if they want to lower their margins or raise the already-high price up to nosebleed levels.
Apple has accepted a 100% price increase for Samsung's LPDDR5X memory, with DRAM supply commitments secured only through the first half of 2026. Tim Cook acknowledged during the Q1 FY2026 earnings call that storage price increases would significantly impact Q2 gross margins.. Apple is evaluating ChangXin Memory Technologies (CXMT) and Yangtze Memory Technologies (YMTC) as new supply sources, attempting to rebuild pricing leverage through supply chain diversification.That's likely only part of the reason. Mac Mini is now "cheap" because everyone exploded in price. RAM and SSD etc have all gone up massively. Not the mention Mac mini is easy out of the box experience.
I considered the mac mini at the time, but the mac mini only makes sense if you need the local processing power or the apple ecosystem integration. It's certainly not cheaper if you just need a small box to make API calls and do minimal local processing.
If you just need "a small box to make API calls and do minimal local processing" you an also just buy a RPI for a fraction of the price of the GMKtec G10.
All 3 serve a different purpose; just because you can buy a slower machine for less doesn't mean the price:performance of the M1 Mac Mini changes.
Sadly not really. The Pi 5 8gb canakit starter set, which feels like a more true price since it's including power supply, MicroSD card, and case, is now $210. The pi5 8gb by itself is $135.
A 16gb pi5 kit, to match just the RAM capacity to say nothing of the difference in storage {size, speed, quality} and networking, is then also an eye watering $300
At that point buy a used macbook air m1.
lol. you need to look at rpi 5 prices again. they are insane.
Do you really need Openclaw now? And not claude code + zapier or Claude code + cron?
That's the point. If you have worse CPU and GPU Windows will be sluggish (it's bloated).
The Mac mini strangely is and has been a very good deal for years now.
This arb you’re talking about doesn’t exist. An m1 studio with 64 gb was $1300 prior to openclaw. You’re not getting that today.
I would have preferred that too since I could Asahi it later. It’s just not cheap any more. The m4 is flat $500 at microcenter.
[1]: https://developer.apple.com/documentation/virtualization/usi...
https://cua.ai/docs/lume https://docs.openclaw.ai/channels/bluebubbles
For the same price in API calls, you could fund AI driven development across a small team for quite a long while.
Whether that remains the case once those models are no longer subsidized, TBD. But as of today the comparison isn't even close.
Assuming, of course, that your legal team signs off on their assurance not to train on or store your data with said Enterprise plans.
With Anthropic you're paying for "more tokens than the free plan" which has no meaning
To be clear, I totally get the idea of running local LLMs for toy reasons. But in a business context the sell on a stack of Mac Pros seems misguided at best.
It is the first local model I've tried which could reason properly. Similar to Gemini 2.5 or sonnet 3.5. I gave it some tools to call , asked claude to order it around, (download quotes, print charts, set up a gnome extension) even claude was sort of impressed that it could get the job done.
Point is, it is really close. It isn't opus 4.5 yet, but very promising given the size. Local is definitely getting there and even without GPUs.
But you're right, I see no reason to spend right now.
The session so far is stored in a file like /tmp/s.json messages array. Claude reads that file, appends its response/query, sends it to the API and reads the response.
I simply wrapped this process in a python script and added tool calling as well. Tools run on the client side. If you have Claude, just paste this in :-)
Suddenly that $40k is quite reasonable because you’ll never pay another dollar for st least 2-3 years.
2-3 years ago people were fantasizing on running local models on a consumer nvidia RTX GPU.
I've been working my way up from a 3090 system and I've been surprised by how underwhelming even the finetunes are for complex coding tasks, once you've worked with Opus. Does it get better? As in, noticeably and not just "hallucinates a few minutes later than usual"?
And while it is stupid slow, you can run models of hard drive or swap space. You wouldn’t do it normally, but it can be done to check an answer in one model versus another.
Sonnet is so fast too. GPT-5.2 needs reasoning tuned up to get tool calling reliable and Qwen3 Coder Next wasn’t close. I haven’t tried Qwen3.5-A3B. Hearing rave reviews though.
If you’re using successfully some model knowing that alone is very helpful to me.
> But I'm def not buying into their rebranding of integrated GPU under the guise of Unified Memory.
But it is Unified Memory? Thanks to Intel iGPU term is tainted for a long time.
Now extrapolating in line with how Sun servers around year 2000 cost a fortune and can be emulated by a 5$ VPS today, Apple is seeing that they can maybe grab the local LLM workloads if they act now with their integrated chip development.
But to grab that, they need developers to rely less on CUDA via Python or have other proper hardware support for those environments, and that won't happen without the hardware being there first and the machines being able to be built with enough memory (refreshing to see Apple support 128gb even if it'll probably bleed you dry).
Basically, too many choices to "focus on" makes non a winner except the incumbent.
https://survey.stackoverflow.co/2025/technology/#1-computer-...
The US 1s? Is that why we have Deepseek and then other non-US open source LLMs catching up rapidly?
World view please. The developer community is not US only.
It wouldn’t surprise me if the deepseek people were primarily using Mac’s. Maybe Alibaba might be using PCs? I’m not sure.
I certainly only use Macs when being project assigned, then there are plenty of developers out there whose job has nothing to do with what Apple offers.
Also while Metal is a very cool API, I rather play with Vulkan, CUDA and DirectX, as do the large majority of game developers.
Only groups of developers more tied to Windows that I can think of are probably embedded people tied due to weird hardware SDK's and Windows Active Directory dependent enterprise people.
Outside of that almost everyone hip seems to want a Mac.
Everyone hip alright, or at least those that would dream to earn a salary big enough to afford Apple taxes.
Remember there are world regions where developers barely make 1 000 euros per month.
People always overlook that CUDA is a polyglot ecosystem, the IDE and graphical debugging experience where one can even single step on GPU code, the libraries ecosystem.
And as of last year, NVidia has started to take Python seriously and now with cuTile based JIT, it is possible to write CUDA kernels in pure Python, not having Python generate C++ code that other tools than ingest.
They are getting ahead of Modular, with Python.
Neural Accelerators (aka NAX) accelerates matmults with tile sizes >= 32. From a very high level perspective, LLM inference has two phases: (chunked) prefill and decode. The former is matmults (GEMM) and the latter is matrix vector mults (GEMV). Neural Accelerators make the former (prefill) faster and have no impact on the latter.
I assume they have a moderate bet on on-device SLMs in addition to other ML models, but not much planned for LLMs, which at that scale, might be good as generalists but very poor at guaranteeing success for each specific minute tasks you want done.
In short: 8gb to store tens of very small and fast purpose-specific models is much better than a single 8gb LLM trying to do everything.
Apple is in the hardware business.
They want you to buy their hardware.
People using Cloud for compute is essentially competitive to their core business.
Remains to be seen how capable it actually is. But they're certainly trying to sell the privacy aspect.
It's the best. We all turned it off. 100% privacy.
So as most people in or adjacent to the AI space know, NVidia gatekeeps their best GPUs with the most memory by making them eye-wateringly expensive. It's a form of market segmentation. So consumer GPUs top out at 16GB (5090 currently) while the best AI GPUs (H200?) is 141GB (I just had to search)? I think the previou sgen was 80GB.
But these GPUs are north of $30k.
Now the Mac Studio tops out currently at 512GB os SHARED memory. That means you can potentially run a much larger model locally without distributing it across machines. Currently that retails at $9500 but that's relatively cheap, in comparison.
But, as it stands now, the best Apple chips have significantly lower memory bandwidth than NVidia GPUs and that really impacts tokens/second.
So I've been waiting to see if Apple will realize this and address it in the next generation of Mac Studios (and, to a lesser extend, Macbook Pros). The H200 seems to be 4.8TB/s. IIRC the 5090 is ~1.8TB/s. The best Apple is (IIRC) 819GB/s on the M3 Ultra.
Apple could really make a dent in NVidia's monopoly here if they address some of these technical limitations.
So I just checked the memory bandwidth of these new chips and it seems like the M5 is 153GB/s, M5 Pro is ~300 and M5 Max is ~600. I was hoping for higher. This isn't a big jump from the M4 generation. I suspect the new Studios will probably barely break 1TB/s. I had been hoping for higher.
5090 has 32GB, and the 4090 and 3090 both have 24GB.
• Studios with Ultra Mx, now 4-way RDMA over Thunderbolt 5, and enormous RAM and SSD options, suggest a strong focus. I don't know what else that RAM would be intended for. Four Studio Ultras (total of 360 GPU cores with M5 Ultras?) with 2TB of unified RAM is a local model beast.
• They refashioned their GPU cores to better support both graphic and neural processing, despite already having focused NPU cores.
I would say they have been leaning into local models for several years.
I expect we will see more models being optimized for smaller sizes, as demand for them increases. With hardware performance and neural focus trending up, and model requirements/quality trending down, the next few years will be interesting times.
What would make me happy: Ultra x 2 (i.e. 2xUltra, 4xMax, 8xPro, 16xM5) packaging in the Studio. With 8-way RDMA. Mac Kong. Perhaps Apple will start making server cards again.
Are they doubling down on local LLMs then?
Apple's stated AI strategy is local where it can and cloud where it needs. So "doubling down"? Probably not. But it fits in their strategy.
I think I'll pass on upgrading.
Honestly, I think that's the move for apple. They do not seem to have any interest in creating a frontier lab/model -- why would they give the capex and how far behind they are.
But open source models (Kimi, Deepseek, Qwen) are getting better and better, and apple makes excellent hardware for local LLMs. How appealing would it be to have your own LLM that knows all your secrets and doesnt serve you ads/slop, versus OpenAI and SCam Altman having all your secrets? I would seriously consider it even if the performance was not quite there. And no need for subscription + cli tool.
I think apple is in the best position to have native AI, versus the competition which end up being edge nodes for the big 4 frontier labs.
Do think it'll be common to see pros purchasing expensive PCs approaching £25k or more if they could run SoTA multi-modal LLMs faster & locally.
"AI" (LLMs) may or may not have a bubble-pop moment, but until it does Apple get to ride it on these press releases and claims. But if the big-pop occurs, then Apple winds up with really fantastic hardware that just happens to be good at AI workloads (as well as general computing).
For example, image classification (e.g. face recognition/photo tagging), ASR+vocoders, image enhancement, OCR, et al, were popular before the current boom, and will likely remain popular after. Even if LLM usage dries up/falls out of vogue, this hardware still offers a significant user benefit.
What is more likely to happen though is that it doesn't take multiple $10B of datacenter and capital to build out models--and the performance against LLM benchmarks starts to max out to the point where throwing more capital at it doesn't make enough of a difference to matter.
Once the costs shrink below $1B then Apple could start building their own models with the $139B in cash and marketable securities that they have--while everyone else has burned through $100B trying to be first.
Of course the problem with this strategy right now is that Siri really, really sucks. They do need to come up with some product improvements now so that they don't get completely lapped.
Being able to Greenfield something new is a tempting pitch to use to poach employees.
And first to market often doesn't win, or else WebVan would still be doing grocery deliveries. We tend to overstate the first-mover advantages because we more easily remember the cases where that turned into lasting dominance while forgetting all the companies that died to first-mover disadvantages.
Unified memory is a serious architectural improvement.
How many GPUs does it take to match the RAM, and make up for the additional communication overhead, of a RAM-maxed Mac? Whatever the answer, it won’t fit in a MacBook Pro’s physical and energy envelopes. Or that of an all-in-one like the Studio.
I just don't get why they're dropping the ball so much on this.
They aren’t dropping the ball, they are being smart and prudent.
I love the push to local llms. But it’s hilarious how apple a few years ago was so reluctant to even mention “AI” in its keynotes and fast forward a couple years they’ve fully embraced it. I mean I like that they embraced it rather than be “different” (stubborn) and stay behind the tech industry. It’s the smart choice. I just think it’s funny.
So yes, the LLM should be inferencing on your prompt, but it should also be inferencing on 25,000 other things … in parallel.
Those are the compute needs.
We just need compute everywhere as fast as possible.
I also haven’t seen any improvements in the frontier models in years, and I’m anxiously awaiting local models to catch up.
That’s makes you so far out at the end of the curve even professionals can’t see you.
This correlation of Apple and privacy needs to rest. They have consistently proven to be otherwise - despite heavily marketing themselves as "privacy-first"
https://www.theguardian.com/technology/2019/jul/26/apple-con...
> To use features such as these, you must enable Location Services on your iPhone and give your permission to each app or website before it can receive location data from Location Services
> By enabling Location Services for your devices, you agree and consent to the transmission, collection, maintenance, processing, and use of your location data and location search queries by Apple and its partners and licensees to provide and improve location-based and road traffic-based products and services.
Android and every other consumer general purpose OS lets you read GPS coordinates from the sensor without telling anyone.
App installs: Any app installed from the App Store obviously tells Apple you installed it. Apple does certificate verification for every side-loaded app, where Apple is the CA. There is no way to install an app on iOS without telling Apple.
Android and every other consumer general purpose OS lets you install apps without telling anyone.
https://andreafortuna.org/2025/11/30/hidden-metadata-reveals...
Do I wish Apple was way more transparent and gave users more control over gatekeeper and other controversial features that erode privacy? Absolutely.