Bad news: Apple drops high-memory Mac Studio configs
Posted by jzn21@reddit | LocalLLaMA | View on Reddit | 136 comments
Looks like Apple has quietly killed off the higher-memory Mac Studio options. The M3 Ultra Mac Studio is now only available with 96GB RAM. The 512GB option was already removed back in March, and now the 256GB config is gone too.
Apple has said both the Mac Studio and Mac mini will stay supply-constrained for the next few months. The Mac mini is also stuck at 48GB RAM max for now. Probably their high-memory chip stock got too expensive to keep producing.
This is a real bummer for us! Big unified memory configs were one of the few (relatively) affordable ways to run large models locally. I am glad I own the M3 Utlra 512, will definitely keep this on (my favorite local model is Qwen 397b atm).
zaphodbeeblebrox00@reddit
This makes the older high memory Mac Studios look a lot more attractive. They were expensive, but for local LLMs the whole appeal was lots of unified memory in one quiet box. A 96GB ceiling changes that a lot.
the-username-is-here@reddit
Just get DGX Spark, or even two.
Virtamancer@reddit
I can’t for the life of me understand why people want a mediocre amount of memory, that’s super slow, for $4k + tax + tip.
BrewBigMoma@reddit
$4,800 now. Prices dun gone up.
the-username-is-here@reddit
Well, good luck running 70B+ models on your superfast consumer GPUs. 😄
Virtamancer@reddit
For my local use cases the maxed out m5 MacBook with 128gb at 600gb/s is large and fast enough for 25% more than this weird ass product.
the-username-is-here@reddit
For me running models which consume most of RAM on Macbook makes the Macbook almost unusable. Not mentioning listening to jet engine on the desk with extra coil whine.
It can be used for inferrence, sure - as long as you don't do anything else on it.
Virtamancer@reddit
Yours might be broken?
I have a maxed out M2 Max and m5 max mbp. The fans cook on the M2 Max when running models, but they’re still quieter than even the fanciest windows machine. You can fill >70gb and it doesn’t slow anything down.
On the m5 max the fans don’t need to spin up as urgently or for as long or at as powerful a speed. And here again, you can fill basically all the ram except what you need for tasks and it has no perceptible affect on the performance of the machine.
the-username-is-here@reddit
Nah, I just have a lot of crap running, so it thrashes quite a bit. It's 14-incher in clamshell, not the best thermals too.
FullOf_Bad_Ideas@reddit
It's small and relatively power efficient, influencers and people with small apartments love those kinds of things more than janky mining rigs.
The_Hardcard@reddit
You don’t understand that when the alternatives cost so much more? All you have to do is price match the alternatives. Not to mention some of the alternatives requiring a hell of a lot more space and energy.
Different people make different tradeoffs, no need for agreement. But surely you realize no one is passing up small compact boxes of 128 - 512 GB of high speed memory that also costs no more than $4 - $6k.
Choperello@reddit
You tip your Amazon driver?
fundthmcalculus@reddit
just the tip.
twack3r@reddit
Slow af memory bandwidth kills tg
Dramatic_Entry_3830@reddit
It's kind of a use case thing. Dgx is much faster in prefill tho. same goes for strix halo. If you use it to read tons of stuff agentic pp is more important in my opinio. (Task time is lower overall)
the-username-is-here@reddit
Exactly. For agentic loop prefill is more important, takes most of the time.
Competitive_Ideal866@reddit
I was sceptical so I checked. Prompt processing is 10.75x more tokens than generation. Prompt processing on this M5 Max running Qwen 3.5 122B A10B Q4_K_XL is 9.3x faster than generation. So prompt processing is taking ~50% of the time for my agentic workloads.
There might be some selection bias here, of course, because I built everything to run on this machine and will have been optimising the bottlenecks.
kweglinski@reddit
caching anyone?
Dramatic_Entry_3830@reddit
Sure if you let it run through every file on your systems and the Internet too...
I don't know what you are doing with your models but I'm waiting for the models to digest files and images and sub agents doing searches most of the time.
kweglinski@reddit
Fair! but do you really just push full context at once and then throw it away in agentic behaviour? or do you incrementally build the knowledge? The problem with PP is most noticable when you don't have caching because with each turn you have to process again everything that was there in context. e.g. for agentic coding looking at my last days omlx stats - 336.4M prefill tokens with 305.3M cached tokens \~90.8% cache efficiency. It's similar with research. Numbers drop a bit if you're hard on spawning sub agents but then you're not carrying as much data along the way (if you consider samescenario) so the PP is not that big of the issue.
Competitive_Ideal866@reddit
I assumed perfect caching.
the-username-is-here@reddit
Yeah, because everyone's use case is benchmaxing TG on empty context.
It has a lot that slow memory and A LOT of compute for prefill.
twack3r@reddit
Slow af prefill kills ttft
BumbleSlob@reddit
M5 fixed this and PP is 4x as fast
Time-Froyo642@reddit
I wonder what the 512gb models will go for on the secondary market
Anbeeld@reddit
Probably because they want to use all this RAM for upcoming M5, that's it.
skrugg@reddit
Would be my guess too
BallsInSufficientSad@reddit
No it's not the same RAM. The reason is that they aren't going to offer the M5 with less RAM than the M3, so they have to remove those M3 listings so people don't continue to demand the M3 versions.
tldr; The M5 Studio is going to have a max config of 96GB RAM. Fucking awful.
I'm so happy I dished out the cash for the 512GB M3. I should have taken out a loan to get 2.
CognitiveDiagonal@reddit
Fuck, I hope that's not true. I was waiting to get a 128GB M4 Max, but now only the 64GB remains available in my country. I hope they make a 128GB M5 Max too, but your reasoning is very sound.
michaelsoft__binbows@reddit
M5 Ultra studio maxing out at 256gb would be reasonable (if disappointing), maxing it at 96 would make zero sense when a macbook already exists with 128.
AlwaysLateToThaParty@reddit
Second hand market should be interesting.
genuinelytrying2help@reddit
>The M5 Studio is going to have a max config of 96GB RAM.
sorry this is just inconceivable to me, like readjust your priors this simply will not happen
Ok-Drawer5245@reddit
Assuming they can get any ram
ego100trique@reddit
You mean they are using it for the Neo lol
RespectableThug@reddit
I know you’re kidding, but those Neos are selling like crazy. I’ll bet Apple generates a TON of lifetime value from those even though they’re cheap. Great way to get people into the ecosystem.
alberto_467@reddit
They're also a foot in the door for all the students and younger people who will probably upgrade to a new (and more premium) mac in the next 5 years. Or that will enjoy the OS and decide to switch to iPhone for the ecosystem.
That's a lot of lifetime value for sure.
classic-engimatic@reddit
You literally just said what the other guy said, just changed some words…
alberto_467@reddit
Jeez you seem fun at parties
classic-engimatic@reddit
I am
Faktafabriken@reddit
In addition to what was said before: Neo really lowers the hurdles to enter the apple ecosystem. In addition to generating profits in itself, it can also drive adoption of other apple products such as iPhone, more expensive laptops in a few years etc
switchbanned@reddit
Not only that, but the income from Neo itself is a drop in the water compared to the lifetime value for apple from driving adoption of other apple products such as iPhone, more expensive laptops in a few years etc. Neo really lowers the hurdles to enter the apple ecosystem in addition to generating profits in itself
fatboy93@reddit
Ok gemma4's thinking loop
Hortos@reddit
From my experience most mac buyers that aren’t using them to actively make money buy once and use them until several years after their model loses support and an app or something stops functioning due to an older macos version. So first gen NEOs are going to be in circulation for a long time.
ego100trique@reddit
They don't even have enough A18 chips left for making more of them which is quite crazy ngl
ithinkilefttheovenon@reddit
Can’t they just make more?
xrvz@reddit
Can't AI data centers just download more RAM?
ego100trique@reddit
They mainly released the neo to get rid of the remaining A18 chips they had
ithinkilefttheovenon@reddit
Ok, but if they run out why wouldn’t they just make more? It’s not like A18 is some rare artifact that they can’t make more of.
869066@reddit
They can and probably will, but they’ll have to restart production lines to make new A18 chips which isn’t nearly as easy as just using surplus chips
More-Curious816@reddit
Lol. Suffering from success. the dream of every company ever
DunderSunder@reddit
People getting trapped in that "ecosystem". I almost feel bad.
ego100trique@reddit
Iphone/mabook interoperability is quite good the rest is quite shit, especially airpod audio switch.
The only one that are able to do that properly is actually Huawei for some reason.
MerePotato@reddit
The thing I can't get over with them is the horrific response time on the display, its like 18ms
AttitudeImportant585@reddit
cant believe no one has mentioned this yet but this was debunked in a recent tim cook tweet
wen_mars@reddit
I'll take the opposite view: They can't get enough memory so they're sacrificing the highest memory models
jharsem@reddit
Precisely, Apple is in the business of shipping units - likely they sacrificed parts of their M3 Ultra line to have memory capacity for their upcoming product line, likely based on M5 - if anything this may tell you about the type of memory you're likely to see.
Durian881@reddit
Hope so too. I'm looking forward to getting a M5 Ultra with at least 256GB RAM.
RemarkableGuidance44@reddit
For as little as $20,000
Beginning-Bug-7964@reddit
It would not surprise me if the "one more thing" on the apple keynote is a 0.
alexp702@reddit
Nvidia: DGX 100k workstation! Apple: hold my beer!
power97992@reddit
A dgx b300 has 8 TB/s of bw and 28 petaflops of fp4 sparse… the mac studio will have 1.26 TB/s of bw and 66 tflops or .066 pflops of fp8( it doesnt support native fp4) , but it will probably have 512-1024 GB of ram
alexp702@reddit
You're assuming its a Mac Studio as we see today. Apple could pack multiple M5U into a package with HBM memory next time round.
power97992@reddit
Hbm is expensive but they are planning cheaper mini hbm for the new in 2027 though and perhaps also for the mac.
Exciting_Garden2535@reddit
Apple silicon chips are SOC, so the memory is actually *inside* the chip. My guess is that they just stopped producing M3 Ultra chips because they're now producing M5 chips, and all the old M3 Ultra chips have sold out.
Baldur-Norddahl@reddit
The memory controller and many other functions of a chipset is inside the chip. But not the actual RAM. That is still external chips. They do solder the RAM, which is why it can't be upgraded.
Exciting_Garden2535@reddit
This is the M3 Ultra board with the big SoC right in the middle. In the second image, you can see the memory right in the middle of SoC. The whole SoC is not one crystal, but still, the memory is a part of the SoC, under the same SoC case, and cannot be outside it.
Baldur-Norddahl@reddit
You might be right, but we can't actually see the LPDDR5 packages on the pictures. The second image is just the backside of the carrier PCB.
gomezer1180@reddit
I hope that’s the case, if they are limiting memory to help the closed cloud models, or they are entering that business model then it’s going to get pretty grim soon for local models.
We’re already getting shafted by the fleet of companies making new models limiting their open source models to single or double digits Billion parameters.
armeg@reddit
I think Apple may be of the opinion that the current state of LLMs is too non differentiated and it’s essentially a commodity market. Apple as a company generally doesn’t want to enter a commodity market unless it can cut its own niche. So they’re sitting back and playing wait and see.
eli_pizza@reddit
Alternately: they just whiffed on it. The LLM features they have shipped have been surprisingly bad.
armeg@reddit
Maybe but we haven’t seen them release a single LLM publicly, just extremely limited uses of AI like summarizing. Even IBM has been releasing models.
ElementNumber6@reddit
So you say they're turning 6,400 MT/s LPDDR5 into 9600 MT/s LPDDR5X? 🤔
photobydanielr@reddit
Duh, you put them in a blender and then pour the mixture into a mold put it in the oven and dehydrated it for four hours, now it takes half as much space with double the nutrient density
falcongsr@reddit
quit divulging trade secrets, you signed an NDA!
JacketHistorical2321@reddit
Exactly. It's not that difficult to recognize. They sold out existing m3 product and are focusing production on m5
FullstackSensei@reddit
Tim Cook seems to have a different opinion in the last earnings call. But what would he know, it's not like he's on reddit.
SwordsAndElectrons@reddit
Well, I wasn't on the earnings call and still don't know what he said, so thanks for the insightful comment.
Lhurgoyf069@reddit
What did he say exactly?
PusheenButtons@reddit
You could tell us what he said in the earnings call maybe, since it’s not in the article…
AykutSek@reddit
Is this because of memory supply or something?
The whole reason I was looking at the Mac Studio was the 256/512GB unified memory configs. 96GB is fine, but not really the same thing for local LLMs. Pretty annoying since I was actually considering the 512GB model.
redditor0xd@reddit
It’s not that great for running massive local LLMs. The prompt tps is what kills it. Anything above 30B active parameters and its slows down considerably especially with KV context full precision or turbo quanted.
doradus_novae@reddit
Second, this, I would take a 96 gig 6000 any day over a Mac m3 with 512 GB memory.
I have both at home and there is no comparison.
The m3 is honestly a giant letdown
mattiasso@reddit
Apple bought RAM in advance. But the shortage is lasting long.
Etroarl55@reddit
2027-2028 all ram is already reserved, people are saying it’s in the air to go well into the 2030-2035.
Zyj@reddit
2040 at minimum. My grand children might be able to buy new RAM.
reasonablyminded@reddit
This is the right move, Mac studios all going for localLLM is a bitch and not what the product was intended for.
Those clients won’t create the ecosystem value that they look for, as in, buying all the other devices, buying stuff on the App Store, Apple one, etc.
xBlaze121@reddit
thank god i didnt wait for the m5 mac mini i probably would have gotten fleeced for 32gb lol
Clean_Hyena7172@reddit
The morbid reality is that someday soon we won't be able to buy any local setup, we'll be forced to rent computer access entirely from Big Tech companies. Fuck this timeline.
uutnt@reddit
You can still buy the memory, but at a much steeper price. And the same applies to big tech. I's not an evil corp conspiracy either. Just a matter of supply/demand and economies of scale. The fact is data-centers get higher utilization, compared to somebody running a single desktop at home.
Clean_Hyena7172@reddit
*You can still buy for now
I know it's just business, but from a business angle they are looking at what will make them the most money and that happens to be catering to enterprise over consumers and selling us computer subscriptions rather than one-off computer purchases.
megadonkeyx@reddit
so very dissapointed [puts imaginary spare $10,000 back in ye olde coin purse]
Kaokien@reddit
Realest comment lmao, watching inventory for a product I won't be buying or need anytime soon
uktexan@reddit
I think you mean “affording”.
Desperate to have one, desperate to find said imaginary $10k in some imaginary coin purse or couch 😂
OldHamburger7923@reddit
At the rate we are going, $10,000 will buy you a 32gb laptop in a few years
sloth_cowboy@reddit
Latest trend, "coaching as the youth call it!" More at 10.
spyboy70@reddit
It better be one of those old school 1970's squeeze ones like this
Pleasant-Shallot-707@reddit
1) there’s a parts crunch
2) they’re probably not manufacturing that model anymore with the M5 Ultra being announced in June so they sold out of their existing stock
OkPack2268@reddit
What’s the story then? I’m new around here
Pleasant-Shallot-707@reddit
There’s no story
mjsxi__@reddit
its bananas — its like they've ceded having thoughts and are now wholly reliant on the ai to tell them everything.
SergeantBeavis@reddit
The Mempocalypse is real. Or the M5 based Mac Studio and MacMini are right around the corner.
awsom82@reddit
Good sign
OkPack2268@reddit
Why?
awsom82@reddit
Memory allocation to newer models
rhythmdev@reddit
they can keep all the ram for themselves, dont care at this point since i went for 3090&5090 setup already...
lorddumpy@reddit
I can't believe I'm typing this but I've been on the hunt for a ~$2500 5090 since release. It's crazy how messed up the market is.
Zakmackraken@reddit
Don’t Apple have RAM on chip? And don’t they make the chips (via TMSC or whoever) so why a shortage?
Minute_Attempt3063@reddit
they are likely making useful for the M5 version.
FullOf_Bad_Ideas@reddit
That sucks, I think it's not likely that we'll see 512GB or even higher memory in Mac Studio M5. That market is probably going away, at best it'll be out but it'll cost more than a few kidneys.
munkiemagik@reddit
No point me personally specualting on something I dont know anything about nor have any impact on, But I sure hope this isnt a forewaring of an indutry-wide precedent being set and AMD for exmaple decide to start restricting the amount of RAM on their APU's eventually.
I was jsut syaing in another thread a few days ago how I would happily dunmp my muliti 3090 machine if/when these AMD APU systems at some point int the future start kicking up their memory bandwidth much higher than current 250GB/s levels
mjsxi__@reddit
nah they just have a new chip theyre diverting the memory too before they launch it
FullOf_Bad_Ideas@reddit
And do you think they'll be selling 512GB variant of that too? What if.. not?
JacketHistorical2321@reddit
Ya... Cause they're saving stock for the next m5 studio refresh. I don't understand how people can't recognize this...
redboy33@reddit
Ah yup.
ohgoditsdoddy@reddit
Just a few days ago there was an article about how Apple weathered the RAM crisis very well. 🤷♂️
falcongsr@reddit
yeah by holding back the tiny amount of inventory they had on hand and letting delivery times stretch out months.
shuozhe@reddit
Saw it yesterday or the day before on Chinese apple site. Wanted to see if I could get one cheaper there..
Common-Membership503@reddit
that is super frustrating honestly. i was really hopin to grab one of those high memory units for running larger models locally cuz my current setup is just struggling with quantized weights. maybe lookin into a used m2 ultra is the way to go for now since those still support the higher ram configs
MrPecunius@reddit
Wait a month. M5 is worth the wait.
rapsoid616@reddit
Are these mac minis really life changing for this space? I am currently doing everything in my windows pc with some safety nets I use and kind of expect them to be stable. I never owned or tried this method so I am genuinely interested.
evia89@reddit
Good for playing with LLM if you need security. I rather buy API sub like opencode/minimax
iamapizza@reddit
They aren't. Just different ways of doing similar things.
Limp_Classroom_2645@reddit
Bad news for apple than for us because aint nobody buying that shit
thisguynextdoor@reddit
They are certainly sold out because of nobody wants them
YoungSuccessful1052@reddit
Also noticed the M4 Max Mac studio is limited at 64GB now instead of 128GB. So with the mini limited at 48GB the only way to get anything usable is the M5 Max Macbook Pro now which tops out at 128GB. I hope The M5 pro mac mini will go back to 64GB and the Mac Studios go back to 128GB for the M5 Max and 256 for the M5 Ultra... (I kind of already gave up on hoping for the 512GB to be back.)
rudidit09@reddit
I hope so, it’s a strange world where MBP now has most ram
StrangeOops@reddit
Was thinking of selling my Mac Mini M4 Pro 64Gb + 2Tb but maybe I’ll hold on a little longer.
tarruda@reddit
Suddenly those $10k for the 512G model are looking so cheap now...
Skystunt@reddit
They can’t offer current and next gen high ram in this ram shortage so they chose to focus on the M5 especially the 100+gb macbooks
srigi@reddit
Let’s hope that’s the case. Because this theory is only speculation or… pure hope.
There is no indication, only hope that 256/512 options re-emerges with M5 Ultra in June/Jule at WWDC.
raysar@reddit
They love money and margin are too low for the high ram configuration ...
evolvingwax@reddit
It’s almost as if a new product line is about to launch and there is a rational reason for this.
Beginning-Bug-7964@reddit
If the inventors of the massively overpriced RAM upgrade grift have decided to forgo those obscene profits in the short term, then the long term plan is most assuredly more obscene.
StupidScaredSquirrel@reddit
People don't see it as a price hike if you just sell new products at a higher price, they do if you up the price of existing stuff. That's my guess anyway
a_beautiful_rhind@reddit
I wonder what the sales of the high memory macs were. If they were a niche, I can see why they'd rather produce more low memory ones to stretch the chips they could get.
RoomyRoots@reddit
Kinda expected. Market is still fucked and new chips should be coming.