"Storage performance is up to twice as fast, with a next-generation SSD architecture built on PCIe Gen 6..."
This is the first personal computer I've noticed that has PCIe Gen 6 storage. I've only seen enterprise PCIe Gen 6 SSDs up until now. Gen 5 SSDs in consumer devices already have high temperatures and thermal throttling, so I'm worried about how Apple's implementation will perform (I know they don't use off-the-shelf SSDs anymore, but I'd imagine the temps would still be a problem).
Im curious, how often is SSD / storage speed performance really useful? I feel like for many people it's akin to gigabyte wifi in that its nice to have, but not really particularly necessary
Due to pricing insanity (not that Apple prices weren’t insane before the ram/ssd shortages) I’m not in the market but I do wonder if my next computer should be a Mac Studio instead of a MBP that lives its life docked. Might be better to just run a Studio and Neo for the very few times I actually need remote capabilities.
I've been thinking about this a fair bit recently.
We make a lot of price/performance compromises for having an attached screen and keyboard on our computer. That was what got me started.
Then I remembered the days of having to go to a special corner of the house to use a computer, vs now when I have a computer with me all the time. In my bag, on the sofa, on the train. Hell, I'm writing this on the work MBP while waiting for an appointment.
And you know what, I think I got more done when I went and sat in a corner of the house all those years ago. I set up an area for "computer work", and it worked really well.
I have a home office, but it's a jumble of cables going into docking stations and all sorts of weird stuff. I think if I streamline it and turn it into a proper "computer room", I might get some of that mojo back. I might even convince my partner that surrendering the home office and having a corner of the den might be good - she can watch TV while I tinker. And I won't be balancing a laptop on my knee and trying to do two things at once.
And the price/performance thing comes back in. Hmm.
There is a definite mental aspect for most WFH folks to having a space that is dedicated to work. I'm not unique in saying this, but the way I put it is "If you work from anywhere in your house, then you're always at work."
And that, from mental load standpoint, is not healthy for most folks.
I've done that, but I chose a Framework Desktop instead. The latest Fedora is closer to Snow Leopard than anything Apple has to offer. Downside is that I still have my MBP because of the lock-in and occasionally pick it up to do computer stuff in weird places.
I got used to the hub-and-spoke model at home (previously thin terminal, server-client, etc.). Big ole desktop/server and smaller devices that (ab)use it remotely. Roam around with a smaller computer/tablet/phone. Tailscale to bind it all together.
If your computing needs line up, it's a very serviceable approach.
Another option is a 'maxed out' mini e.g M5 Pro/64GB. Mini is super easy to travel with if you know there'll be a 'dock' at the far end e.g office<>home or whatever.
I had the same view recently. For the past year I’ve been using my iPad to remote into both my MacBook Pro M1 Max and my racked Linux workstation at home using Jump and moonlight/sunshine, respectively. Never looked back
Having replaced my MacBook with a Mac Mini, I would reconsider. The MacBook is just such a _complete_ package. Great speakers, great keyboard, fantastic screen, the fingerprint sensor thingy.. Takes a lot of gear to match that
I’ve had iMacs for almost 20 years. My last one was indeed my last. Without target display mode (use the Mac as a monitor), I’m ditching a perfectly good monitor. I was going to buy a Mac Studio and a good monitor to replace the iMac until the spouse reminded me that we are now retired and will spend time in a camper. So a MBP for me (and BenQ’s Mac-specific monitor), but others might do well to consider a Mac Mini/Studio.
iMacs are great for a lot of use cases, but my image of the typical HN user would prefer to keep the monitor separate.
My M1 Pro drives 4 screens no problem. The trick is that you can only attach two through one Thunderbolt port, so you have to attach the 3rd one via the other port instead of your dock. (The 4th screen is the laptop's built in screen)
At my current job, one of my biggest blunders was thinking "Let me just order the same hardware as most of my teammates to avoid unnecessary complications".
Most, if not all, of our current work happens on remote cloud vms. Now I'm stuck with carrying a 3KG monstrosity to work.Every day.
Absolutely no positives compared to my ThinkPad that weighed less than half in my previous job.
Remote development is "good enough" these days. With VS Code, Development containers etc... Having a light weight, portable laptop is so much nicer than a laptop that you can't even rest on your lap for long duration.
The only thing you need to be mindful of using light weight laptops is not having enough RAM to fit all your browser tabs..
Yeah I stepped down to a smaller MBP. the only reason I didn't go for an air is my home setup with dual monitors involves using the HDMI port. There probably is a dock or something out there though.
I totally agree I use containers through propelcode.app which means I can code on any device including my phone and the environment is the same as the server target. It’s even more important with agents that can delete files and change OS settings. I never need to worry that a coding session ends up breaking my main OS environment by deleting or changing a file.
I dusted of my lightest computer with an M1 chip and use Tailscale to make my network virtual from anywhere. Been running a. Pi5 as a main house hub and an M1 Pro as an always on Mac. It would be nice to go all out and make a Studio a hub I can just screen share into for major compute.
I had the same thought, I grabbed a studio two years ago for this reason and it’s been great. 99% of the time lack of portability isn’t a concern. Every now and then (e.g. travel) I notice the limitation, but it’s not much of an inconvenience to just not do some work for a bit.
Plus remote work is getting easier and easier. There are so few instances when I'm not able to get online. If we lived in a world where hardware were getting cheaper, it might make sense to splurge. In this environment I think the Neo is perfect.
Build quality of the Neo is extremely good, I love the keyboard — it’s more tactile and reminds me of early 2010s MacBooks.
I’ll be selling my M4 MBA soon, I genuinely use the Neo more. Huge difference in typing experience.
Great repairability is a plus. It was super easy, and actually fun to open. Felt like unboxing an Apple product. Applied the thermal paste mod for $10 which works excellently; I’ve had it shortly after launch.
And I love the notchless display, even if I wished the color gamut was a bit better.
>I do wonder if my next computer should be a Mac Studio instead of a MBP that lives its life docked.
Having owned 3 MacBook Pros since 2008, the decision to make my next computer be a Mac Studio came down to (1) MacBook thermal throttling that slows down CPUs when it starts to overheat and (2) easier upgrade of Mac Studio SSD with after-market storage module whereas the MacBook requires more complicated disassembly and hot air gun to dislodge the surface mounted SSDs.
I have a brand new M5 Pro MacBook Pro I don't like it when the fans turn on. The Mac Studio will be faster and quieter for the same workloads.
Apple prices have always been insane, there is a reason why during the days it almost went bankrupt, in Europe it could not rival with PC, Amiga, Atari, Acorn.
Recently took a Minisforum 7840hs PC out of rotation as a media PC and made it a full time coding workstation with Proxmox. I do a VM per project due to the nature of agentic editors.
I was using a VM setup on my MBP but it felt like a huge waste, having to leave a laptop on 24/7 when all it did was run Claude Code inside VMs.
I likely will stick with a Macbook Air 15" for next purchase, and beef up my "Claude Server" down the road.
I've been holding out, because I think my next purchase will be a Studio with an Ultra Chip in it. I'm wanting it to be a "forever" server, so I'm holding out while I can.
Supply rumours are next year we see an M7 AI-focused chip with large inference performance upgrades. It's unlikely we'll see heavy upgrades in other areas. If you care about AI, it's worth waiting. If you don't, pull the trigger now. RAM constraints are likely to get worse next year. Or wait 2-3 years and prices should be back to Earth (plus newer and even better chips).
Yeah, but how many years until 128GB+ is attainable by mere mortals again? My 2021 home server build was 64GB of RAM. My 2025 build was 32GB and zram. :/
I just wait for a cycle or two where the leaps and bounds are more like hops and steps. So if the M7 Ultra improves inference by 2x over the M5, but the M9 Ultra only improves by 1.2x over the M7, that's my signal to buy. Unfortunately they haven't slowed down yet.
Yep, this is my exact thought. The pendulum has swung back toward a desktop making more sense for me than a laptop. It all depends on whether there is anything useful to do with an amount of computation that can't be fit into a laptop package. For a long time there wasn't, now there is.
That’s nearly what I do but on a smaller scale. My iPad Pro serves as my laptop 90% of the time, and the 10% of the time I need to actually code and test in a chromium browser I remote into a mini.
I’m doing essentially this, and got a MacBook Neo.
Fir kinda the first time in my life I don’t really have development tools on my personal laptop. I ghostty and openvpn client installed.
I have a large remote linux workstation (2x 8c/16t xeon cpus, 256gb ram, 2x8tb spinning rust disk) and i have my tools over there (along with some VMs).
It works surprisingly well.
Also, the macbook neo is a surprisingly capable little machine.
Because who hates themselves that much? It's the thing I touch and interact with. That's exactly the part that needs to be sturdy, smooth, and pretty. It's the facade to the beast at the other end.
For a desktop you may find yourself better on a Linux or Windows machine price/performance wise.
I personally own an M3 ultra, an M1 max as laptops, but my desktop is a Ryzen desktop I built in 2022 and it was a third in price of the ultra for more power.
I was in the same situation, I used maxed out 15'' M3 Max MacBook Pro docked to Studio Display closed on vertical stand behind the screen. It was fine for office work, but running local LLMs would definitely overheat it. The battery started degrading purely due to heat issues. And it was audible as well.
I decided to get Mac Studio M4 Max, also all maxed out config and the cooling is so much better that I can run local LLMs like Gemma 3/4, gpt-oss 120b all day long without any heat issues or any audible fan noise. So for my use case it was the right decision. I subsequently added 15'' M5 Max MacBook Pro all maxed out to my collection and even though it is slightly faster on LLM inference (I get 100 tokens/s with Gemma 4 27b model), you just can't run LLMs longer than a few minutes. It starts overheating and gets really loud.
Weird, I’ve run LLM batch sessions for hours on my Max M3 MBP. It doesn’t get very loud, though I’m not getting anything close to 100 tok/s on a 27b model, I use a 35b MoE model just to get 90 tok/s. The fan comes on but thermally it never overheats. I do have it in a vertical closed position, though.
That’s what I have been doing for years, it remains in the house secured while I ssh into it from an old thinkpad. You can get air to pair it with it if you really wanna have that seamless flow, otherwise, ssh works well.
I'd never go x86 again after owning an m1. I'd have replaced an x86 laptop 2 or 3 times by now (2020 m1). My last $3500 dell xps 13z before buying the m1 was absolutely horrible.
Laptops can't do agentic engineering. They get hot as hell and battery drains instantly. I think this will promote a switch to desktops for the next couple of years, until we have new mobile chips.
10 grand for 256GB memory. Likely double that for 512GB, but won't be available or finalized until October. Thunderbolt 5 is highest bandwidth external IO available at 120Gb/s. 1.2TB/s claimed max internal memory bandwidth.
Not exactly "future proof" for >1T parameter models but good for targeting specific lower-parameter models, or if you can rely on pipeline parallelism and run a cluster.
About 20 years ago my dad bought me a $5k computer, it was future proof for about "5 years" before we had to upgrade its internal parts (more memory, new graphics card).
It was future proof but not really because it struggled a lot in its final years.
Upgradeable components however could go a loooong stretch towards that goal. It can't be that hard to follow a common form factor for at least the housing across two or three generations to allow a reuse of everything but the main PCB.
Think it would have same memory bandwidth if the RAM was upgradeable?
Would be nice if someone knowledgeable about electrical engineering and manufacturing processes could lay out some valid reasons for manufacturers to integrate RAM onto the motherboard.
It isn't integrated into the motherboard, it's integrated onto the same package as the CPU/GPU which allows for better signal integrity and higher speeds. You can get somewhat close to the same speeds while modular with tech like LPCAMM2, but there are some pretty difficult challenges to overcome to close the gap completely. As an example, the Framework Laptop 13 Pro CPUs support up to 9600MT/s (same as M5), and Micron sells LPCAMM2 modules that can run at 8533MT/s, but the 13 Pro only officially supports 7467MT/s.
If it was upgradable, then yes, spending more on top of it every year would make it future proof, but that's not the point. It's that spending 10 grand doesn't get you a future proof computer today.
> The way the Apple M-series does ram that might be difficult to pull off.
Well it might be an idea to keep the layout of the mainboard and connectors the same.
That way, instead of having to upgrade the whole machine, all it would need is a new mainboard. Framework for example managed to pull that off, and in mobile at that, where constraints are much worse than for a desktop computer.
> Framework for example managed to pull that off, and in mobile at that, where constraints are much worse than for a desktop computer.
It's not the same thing though. On the M-series, CPU and GPU share a unified memory architecture and ram is much more tightly coupled to get it to go faster. A closer example would be the Framework desktop, actually, where memory is also soldered in for the same reason.
The relevant comparison isn't one mac studio to one RTX 6000, it's a 24 channel DDR5 system, which also has ~1.2TB/s of memory bandwidth (or more when Xeon 6 compatible 8800mt/s memory becomes widely available), vastly higher prefill due to more CPU horsepower, orders of magnitude faster networking, can hook into GPU accelerators, can be upgraded etc. A baseline 384GB system from eg Puget is ~30K vs ~12K for the 256GB Mac Studio and you do get value for the money.
It can be more than "slightly", particularly if the model you're interested in (or will be interested in in 6 months) doesn't fit on the mac studio. You also need to account for eg storing 10TB of random checkpoints, load time when experimenting, and so on. When you start actually needing throughput these are all capability gaps in practical use, not just x% benchmark differences.
If you just want to run Qwen 3.8 27B and Deepseek v4 Flash in perpetuity and that's it, there are a lot of solutions that will work and this is a fairly user friendly one.
How's the compute side now, I wonder? Because while the Ultras have impressive memory bandwidth for inference, processing prompts still takes a dog's age on my M3 Ultra. I heard the M5 makes some strides forward in this area, though, and the M7 in particular promises to go a lot further.
M5 is excellent, they’ve finally gotten their own tensor cores.
Good for inference; however if you like to train, data format support and effective performance is limited (M5 Pro). Some hardware features are not exposed or extremely slow.
You’ll be fine for inference, but pales in comparison to what a RTX 6000 Pro can do for compute/matmuls/training.
Except the RTX 6000 will run circles around the Mac studio in just about every way. Memory bandwidth is literally the only spec where Apple is competitive, and while high memory bandwidth is necessary for LLMs to perform well, many people strangely don't understand that memory bandwidth alone is not sufficient.
It's unclear to me how bandwidth scales with multiple connections. Many-to-many does not seem ideal. Daisy chaining would be fine for straight pipeline work. There doesn't seem to be an equivalent of a ethernet switch for thunderbolt 5 though.
Sorry for being lazy, but is there a rough breakdown like "You get sonnet level for M5 and Opus for M5 pro, etc.", or is it still speculative. Or put simpler, do you get Opus level for the 256GB M5 Max?
For local LLMs with a Mac, rule of thumb is you always want an Ultra (due to memory bandwidth). Even an M1 Ultra is superior to an M6 Pro in this regard.
There are no configurations even close to running something comparable to frontier model variants, they're simply far too large, but something like full precision Qwen 35b or DeepSeek 70b at 50+ t/s is well within available configuration, and potential for plenty of room for large context sizes.
1.2 TB/s bandwidth of M5 Ultra comes from two dies of M5 Max (each 614 GB/s) connected together using 4.4 TB/s inter-die fabric.
For a non-quantized Deepseek V4 flash on an ultra, I would estimate about 1000+ tokens per second prefill and 50+ tokens per second on generation. This is actually quite usable and near parity to cloud.
They mention "adds the GPU Neural Accelerators." which, if exploitable for LLM loads, would probably help the prefill a lot
It looks like speculation that Apple would raise the base chip’s maximum RAM from 32GB to 48GB was wrong.
Apple also launched the base M6 today with a 32GB RAM limit, suggesting 512GB may remain the maximum for Ultra chips for some time. Since these Ultra chips combine 16 base chips:
They probably literally don't have enough NAND to go around. 768GB of memory (48GB x 16) is enough for nearly 100 iPhone 17s; that's $800k of iPhones at MSRP, although likely much lower margins than these high-RAM boxes.
Mac Studio with M5 Max starts at $2,499 (U.S.) and $2,299 (U.S.) for education. Additional configure-to-order options are available at apple.com/mac-studio. Mac Studio with M5 Ultra starts at $5,499 (U.S.) and $5,099 (U.S.).
I am in Europe, and the Mac Studio M5 Ultra GPU 64 cores with 96GB RAM is up to 6.649,00 €. Ouch.
To put this into a perspective, Google helpfully reminds:
> A fully configured IBM Personal Computer AT (Model 5170) with expanded memory and storage cost around $5,795 to $6,000 at its launch in August 1984, which equals roughly $18,600 to $19,300 in 2026 USD.
John Dvorak said many, many decades ago (80s/90s) that the computer you want will always cost $3000. That statement has been more/less true for some time periods than others, but with some wiggle room I’ve found it to be accurate enough.
Care to guess the approximate price of the MBP I bought earlier this year?
My heavily-upgraded M1 Max came in slightly over that when I got it five years ago. (Still going very, very strong.)
This new Studio? Can't find a config under $5k I'd bother with. But for the MBPs that number still mostly tracks for the average Pro user. (I buy large and run it into the ground so long I mistake the ground for the computer's remains.)
They still are, if what you want is roughly the same as the prior generations capability with some uplift (making then number up, but say 20% faster or more ram or whatever).
What is changing is that there genuine demand for more capabilities disproportionate to the cost decrease curve. Fab demand and supply constraints have slowed or even reversed some cost decreases - but that is still getting absorbed by the overall systems costs when you are looking at things like laptops. If you all you want is the last decades demand to browse the web and use office - things are cheaper than ever.
> M6 supports up to 32GB of unified memory to multitask across demanding apps and run LLMs on device for secure and private agentic tasks. It also provides up to 170GB/s of unified memory bandwidth — a 10 percent increase over M5 and a 2.5x increase over M1.
Same reason they cut the big options on the existing models, this way they can sell more devices. The additional cost for the additional 512GB would have to make up for the loss of another sold device otherwise. No idea if there would really be that many people buying this then while on the other hand AI stuff makes people do crazy stuff, so...yeah :)
They seem to be suffering from the supply constraints like everyone else. They phased out the higher capacities on the M3 Ultra Mac Studio a while ago, and if you order a 128GB MBP, say, you're looking at six weeks or more for delivery.
They have no fab for CPUs, they are manufactured by Samsung and TSMC. The bottleneck is in manufacturing RAM not CPUs so there is nothing Apple can do here.
True, but you can link them up over thunderbolt or Ethernet. If your goal is to run local LLMs, not all weights need to live on the same computer. You can segment the workload by layers and pass the activations along the lower bandwidth interconnect with a small perf penalty. Also you get double the CPU/GPU cores allowing for better multi user/agent performance.
Neither is the right alternative to compare to. You aren’t going to hit 100% utilization (if you are, ignore me, this doesn’t some to you, and write a blogpost for me to read and share).
The comparison should be against renting in the cloud for the duration of your task for training and research or using pay-per-api-call providers for general inference instead of buying your own hardware (and paying the electricity and cooling bills on top), because let’s face it, the models you want to use are probably the same ones available on inference providers (but, yes, some are more trustworthy than others).
Speaking as someone that does ML/AI research, you are essentially paying a huge premium for being able to just run your Python script at any time without setting up a deployment script and harness to run the job remotely, while your hardware sits essentially idle the rest of the time.
The only way to make the math work is if you rent your hardware in the background for inference while you’re not using it in anger, but despite all the startups and promises that has never become as streamlined as mining bitcoins or shitcoins used to be and they don’t pay out as much as they say they would. Renting your hardware for training is another option but doing that is a lot more involved, options are fewer and farther in between, you won’t get as much utilization out of it, and doesn’t let you feasibly abort running tasks at a moment’s notice.
My card (RTX Pro 6000) is always doing something all the time from my queue; like some synthetic dataset generation up next. I still actively use runpods and openrouter for scaled stuff, I was spending a bit and then did the maths, and invested in it.
The maths to me was basically equivalent to prepaying for 242 days of runpod pricing for the same GPU; and I reckon I'd be able to get 6+ years of use out of this card with 96GB.
Plus there's the resell value -- it's actually appreciated by ~50% since I bought it.
Plus I do really enjoy that it's 100% local. I wouldn't feel comfortable giving my agents this much information if inference wasn't 100% local.
I wouldn't get another one, I wouldn't have as much value, but one is definitely paying off for me on the financial side.
To experiment, its still al ot cheaper to prepare everything locally and then just rent a GPU Node on all of these non hyperscalers.
1-2$ / hour.
I'm not regretting my setup at home as it got paid by my company which makes sense here, but paying for electricity is quite high and makes already 0.3$/hour alone.
I would argue, the most interesting use case for running it at home is some personal agent which you want to run 24/7.
Big ole pool of very fast ram that can be accessed by the CPU and GPU. Lets you run larger models. AMD does the same thing with Strix Halo. I have a 128gb machine at home, and have had difficulties running 120b models, but 70b and below run pretty well.
Depends on your perspective. People think nothing of spending 50-100K on a car that basically gets them to work. But the thing they use for day to day work then gets the evil eye when it costs more than 1K. It's slightly irrational. Not everybody needs a high end mac. But when you do, it sure is nice that you can get one.
I don't actually own a car and my startup is bootstrapped and our salaries are modest. But the one thing we spend on is laptops. I have M4 max pro with 48GB. That thing was on the expensive side (~4.5Kish). But it delivers a lot of value and I spend most hours I'm awake using it. I like fast builds. I like that I can try out open source AI models. And I like just having the option to run those.
We actually lease them and mine costs something like 105 euro/month. Including Apple Care. I don't need a Mac Studio but I could see some roles where that would not be a crazy expense. Even the tricked out version that basically only costs the same as a very modest car.
Boy, oh boy Apple is the new shovel seller during AI gold rush.
Most people at Apple have already realized that their processors are already too powerful for regular users - heck, as a developer my M2 Pro with 32 GB RAM is more than enough for me.
Regular users don’t care about local AI either. So, they will probably extract as much money as possible during AI gold rush, but then we will most likely see Apple
a. Making their software worse (god forbid, forced updates)
b. Making their hardware impossible to repair (as they almost accomplished this already) and easier to break.
I dunno, Apple's been throwing many bones to their customers who use their Macs for AI stuff; like Apple working with, and signing TinyGPU's NVIDIA eGPU drivers.
Plus introducing features like RDMA over thunderbolt, which is critical for distributed inference/training/etc. On the software side, Apple is investing heaps.
It's still ridiculous they don't support expandable NVMes, but the memory being soldered makes sense, you need it for 1.2TB/s bandwidth.
They are still selling high-margin hardware. Apple loves selling high-margin hardware.
Wow, didn't know this exists. So instead of ditching my M1 Pro Macbook for a newer one with more RAM and power I could do this instead (if reasonable regarding pricing).
I wish they were offering 1TB of Unified Memory for the M5 Ultra. I already have an M5 Max MBP w/ 128GB of RAM for running local models, and while there's a /few/ models that I can run in 512GB that I can't run in 128GB that are interesting, where things really shift is at 1TB of memory which allows you run >1T parameter models w/ 4 bit quants reliably. 512GB is just on the edge of "enough", which is maybe the point of maximum frustration considering current memory prices.
Personally, I can't justify dropping the dosh for a 512GB M5 Ultra, but I would be able to justify it to myself if I could get 1TB of memory, because it'd guarantee the flexibility with local models I currently am missing. Seems a huge miss to not offer this... for a price.
I'd consider a 1tb machine at 20k, but I'm not going to pick up a 256gb one at all. 1TB fits a frontier-ish model in memory without massive quantization, which is a very interesting capability for a non-rack piece of compute.
More likely double that, even. I think you'd still see many buyers there. You can spend like $16k alone on a RTX 6000 PRO with a mere 96GB of VRAM now..
I would probably spend up to $30k if I could get 1TB of Unified Memory, because it would allow me a guarantee to run pretty much any local model I want, including >1T parameter models with reasonable quants. I wouldn't be surprised if 512GB is close to $20k when it becomes orderable in October. The justification is less about absolute price and more about price to what it enables. 512GB really doesn't enable much over 128GB for me, but 1TB would massively change things.
I have a bit of paranoia/anxiety about AI, but it's not what most people are concerned with. I understand the limits of these tools very well, and still find them extremely useful. What concerns me is that it's going to become difficult to impossible in the future to run local models which have near-SOTA capabilities in a way in which you can exercise full control of the model. I see the writing on the wall, and its more than worth it for me to invest early to ensure my own capabilities. I am very much not a fan of our "you'll own nothing and be happy" directionality for the world, and I am (at least currently) privileged to have the means to slow that decline for my own self.
RDMA is buggy and Thunderbolt only delivers 1/10th the throughput of native connectivity. 1TB of Unified Memory w/ 1.2TB/s of bandwidth with marginally ~$30k cost is a different story than 1TB of sorta Unified Memory w/ an effective 120GB/s of bandwidth with a marginally ~$40k cost + all the RDMA bugs.
anyone know if this is "pre-order" is coming late October, or "will be available to ship in late October, thus pre-order will be available earlier than that"
Am I the only one that now finds press releases like this similar to "AI Slop"
I know there's tons of marketing language, buzz words and attempts at convincing me of some agenda that isn't super clear without lots of effort in "validating" the slop. I guess its not bad "slop" though if a human put in effort in editing it (imo >50% human curating = not really bad ai slop)
Though I still would prefer I could just get the prompt. What human thoughts, direction and "prompt" went into writing this article? in the same way as we ask for the prompt for AI generated outputs, I would prefer it for human generated output too. For writing at the least. I could have saved time, got the purity of the argument, and got more clear information. I wonder if we can get a future where humans just express their intent with each other and stop trying to hide our agenda; I want a world we can trust each other greater and interpret and act on our goals without the noise of trying to impress or market to each other & the additional words that go into that.
The thing that intrigues me the most is this:
"Storage performance is up to twice as fast, with a next-generation SSD architecture built on PCIe Gen 6..."
This is the first personal computer I've noticed that has PCIe Gen 6 storage. I've only seen enterprise PCIe Gen 6 SSDs up until now. Gen 5 SSDs in consumer devices already have high temperatures and thermal throttling, so I'm worried about how Apple's implementation will perform (I know they don't use off-the-shelf SSDs anymore, but I'd imagine the temps would still be a problem).
Im curious, how often is SSD / storage speed performance really useful? I feel like for many people it's akin to gigabyte wifi in that its nice to have, but not really particularly necessary
Due to pricing insanity (not that Apple prices weren’t insane before the ram/ssd shortages) I’m not in the market but I do wonder if my next computer should be a Mac Studio instead of a MBP that lives its life docked. Might be better to just run a Studio and Neo for the very few times I actually need remote capabilities.
I've been thinking about this a fair bit recently.
We make a lot of price/performance compromises for having an attached screen and keyboard on our computer. That was what got me started.
Then I remembered the days of having to go to a special corner of the house to use a computer, vs now when I have a computer with me all the time. In my bag, on the sofa, on the train. Hell, I'm writing this on the work MBP while waiting for an appointment.
And you know what, I think I got more done when I went and sat in a corner of the house all those years ago. I set up an area for "computer work", and it worked really well.
I have a home office, but it's a jumble of cables going into docking stations and all sorts of weird stuff. I think if I streamline it and turn it into a proper "computer room", I might get some of that mojo back. I might even convince my partner that surrendering the home office and having a corner of the den might be good - she can watch TV while I tinker. And I won't be balancing a laptop on my knee and trying to do two things at once.
And the price/performance thing comes back in. Hmm.
There is a definite mental aspect for most WFH folks to having a space that is dedicated to work. I'm not unique in saying this, but the way I put it is "If you work from anywhere in your house, then you're always at work."
And that, from mental load standpoint, is not healthy for most folks.
I've done that, but I chose a Framework Desktop instead. The latest Fedora is closer to Snow Leopard than anything Apple has to offer. Downside is that I still have my MBP because of the lock-in and occasionally pick it up to do computer stuff in weird places.
I got used to the hub-and-spoke model at home (previously thin terminal, server-client, etc.). Big ole desktop/server and smaller devices that (ab)use it remotely. Roam around with a smaller computer/tablet/phone. Tailscale to bind it all together.
If your computing needs line up, it's a very serviceable approach.
BRING BACK THE COMPUTER ROOM
Another option is a 'maxed out' mini e.g M5 Pro/64GB. Mini is super easy to travel with if you know there'll be a 'dock' at the far end e.g office<>home or whatever.
I had the same view recently. For the past year I’ve been using my iPad to remote into both my MacBook Pro M1 Max and my racked Linux workstation at home using Jump and moonlight/sunshine, respectively. Never looked back
Having replaced my MacBook with a Mac Mini, I would reconsider. The MacBook is just such a _complete_ package. Great speakers, great keyboard, fantastic screen, the fingerprint sensor thingy.. Takes a lot of gear to match that
If your computer never leaves your desk, the iMac is pretty competitive with those features. Even comes with a TouchID keyboard.
I’ve had iMacs for almost 20 years. My last one was indeed my last. Without target display mode (use the Mac as a monitor), I’m ditching a perfectly good monitor. I was going to buy a Mac Studio and a good monitor to replace the iMac until the spouse reminded me that we are now retired and will spend time in a camper. So a MBP for me (and BenQ’s Mac-specific monitor), but others might do well to consider a Mac Mini/Studio.
iMacs are great for a lot of use cases, but my image of the typical HN user would prefer to keep the monitor separate.
Not an option if you're trying to drive 3 decent screens.
My M1 Pro drives 4 screens no problem. The trick is that you can only attach two through one Thunderbolt port, so you have to attach the 3rd one via the other port instead of your dock. (The 4th screen is the laptop's built in screen)
Macbook (base) yes, but my MBP (M3 Max) is driving 4 screens for me without issue.
At my current job, one of my biggest blunders was thinking "Let me just order the same hardware as most of my teammates to avoid unnecessary complications".
Most, if not all, of our current work happens on remote cloud vms. Now I'm stuck with carrying a 3KG monstrosity to work.Every day.
Absolutely no positives compared to my ThinkPad that weighed less than half in my previous job.
Remote development is "good enough" these days. With VS Code, Development containers etc... Having a light weight, portable laptop is so much nicer than a laptop that you can't even rest on your lap for long duration.
The only thing you need to be mindful of using light weight laptops is not having enough RAM to fit all your browser tabs..
Yeah I stepped down to a smaller MBP. the only reason I didn't go for an air is my home setup with dual monitors involves using the HDMI port. There probably is a dock or something out there though.
I totally agree I use containers through propelcode.app which means I can code on any device including my phone and the environment is the same as the server target. It’s even more important with agents that can delete files and change OS settings. I never need to worry that a coding session ends up breaking my main OS environment by deleting or changing a file.
I dusted of my lightest computer with an M1 chip and use Tailscale to make my network virtual from anywhere. Been running a. Pi5 as a main house hub and an M1 Pro as an always on Mac. It would be nice to go all out and make a Studio a hub I can just screen share into for major compute.
I had the same thought, I grabbed a studio two years ago for this reason and it’s been great. 99% of the time lack of portability isn’t a concern. Every now and then (e.g. travel) I notice the limitation, but it’s not much of an inconvenience to just not do some work for a bit.
Plus remote work is getting easier and easier. There are so few instances when I'm not able to get online. If we lived in a world where hardware were getting cheaper, it might make sense to splurge. In this environment I think the Neo is perfect.
Build quality of the Neo is extremely good, I love the keyboard — it’s more tactile and reminds me of early 2010s MacBooks.
I’ll be selling my M4 MBA soon, I genuinely use the Neo more. Huge difference in typing experience.
Great repairability is a plus. It was super easy, and actually fun to open. Felt like unboxing an Apple product. Applied the thermal paste mod for $10 which works excellently; I’ve had it shortly after launch.
And I love the notchless display, even if I wished the color gamut was a bit better.
This is my setup (except I have an Air because the Neo didn't exist yet). And it's great. Tailscale makes it trivial.
So, do you use VNC (or ssh) to access the Mac Studio at home? Or do you only access other stuff that's living in your home network?
not OP but tailscale ssh does the trick for me.
I have been thinking a lot about buying a very fast desktop and a cheap laptop that uses remote desktop to connect to the fast computer.
But in the end I ended op buying a Lenovo Legion and put Linux on it.
Laptops are so fast these days that I didn't want to be bothered with setting up connectivity to a remote desktop.
But if your laptop never leaves your desk I think a desktop computer is a great option. Relatively cheaper and easier to maintain and upgrade.
>I do wonder if my next computer should be a Mac Studio instead of a MBP that lives its life docked.
Having owned 3 MacBook Pros since 2008, the decision to make my next computer be a Mac Studio came down to (1) MacBook thermal throttling that slows down CPUs when it starts to overheat and (2) easier upgrade of Mac Studio SSD with after-market storage module whereas the MacBook requires more complicated disassembly and hot air gun to dislodge the surface mounted SSDs.
I have a brand new M5 Pro MacBook Pro I don't like it when the fans turn on. The Mac Studio will be faster and quieter for the same workloads.
Apple prices have always been insane, there is a reason why during the days it almost went bankrupt, in Europe it could not rival with PC, Amiga, Atari, Acorn.
Recently took a Minisforum 7840hs PC out of rotation as a media PC and made it a full time coding workstation with Proxmox. I do a VM per project due to the nature of agentic editors.
I was using a VM setup on my MBP but it felt like a huge waste, having to leave a laptop on 24/7 when all it did was run Claude Code inside VMs.
I likely will stick with a Macbook Air 15" for next purchase, and beef up my "Claude Server" down the road.
I've been holding out, because I think my next purchase will be a Studio with an Ultra Chip in it. I'm wanting it to be a "forever" server, so I'm holding out while I can.
Supply rumours are next year we see an M7 AI-focused chip with large inference performance upgrades. It's unlikely we'll see heavy upgrades in other areas. If you care about AI, it's worth waiting. If you don't, pull the trigger now. RAM constraints are likely to get worse next year. Or wait 2-3 years and prices should be back to Earth (plus newer and even better chips).
I'm waiting this out.
Yeah, but how many years until 128GB+ is attainable by mere mortals again? My 2021 home server build was 64GB of RAM. My 2025 build was 32GB and zram. :/
Then you'll always be waiting, there's always something new the industry tries to tempt you with.
I just wait for a cycle or two where the leaps and bounds are more like hops and steps. So if the M7 Ultra improves inference by 2x over the M5, but the M9 Ultra only improves by 1.2x over the M7, that's my signal to buy. Unfortunately they haven't slowed down yet.
Buy in 2 years, or buy now and have it last 2 years less than forever.
Yep, this is my exact thought. The pendulum has swung back toward a desktop making more sense for me than a laptop. It all depends on whether there is anything useful to do with an amount of computation that can't be fit into a laptop package. For a long time there wasn't, now there is.
That’s nearly what I do but on a smaller scale. My iPad Pro serves as my laptop 90% of the time, and the 10% of the time I need to actually code and test in a chromium browser I remote into a mini.
I’m doing essentially this, and got a MacBook Neo.
Fir kinda the first time in my life I don’t really have development tools on my personal laptop. I ghostty and openvpn client installed.
I have a large remote linux workstation (2x 8c/16t xeon cpus, 256gb ram, 2x8tb spinning rust disk) and i have my tools over there (along with some VMs).
It works surprisingly well.
Also, the macbook neo is a surprisingly capable little machine.
If all you want to do is remote into your desktop, Neo seems like overkill. Why not just get a $200 Chromebook and save yourself $500?
Because who hates themselves that much? It's the thing I touch and interact with. That's exactly the part that needs to be sturdy, smooth, and pretty. It's the facade to the beast at the other end.
Because the screen, keyboard, and especially trackpad on a $200 Chromebook sucks?
Also: the enclosure, and the hinges.
How about business where you send in all your old devices and get back a SSD using using their memories
For a desktop you may find yourself better on a Linux or Windows machine price/performance wise.
I personally own an M3 ultra, an M1 max as laptops, but my desktop is a Ryzen desktop I built in 2022 and it was a third in price of the ultra for more power.
Do it! I went with the Mini/Neo combo. I don't need MBP power when out and about. When at home, the Mini is all I use.
I was in the same situation, I used maxed out 15'' M3 Max MacBook Pro docked to Studio Display closed on vertical stand behind the screen. It was fine for office work, but running local LLMs would definitely overheat it. The battery started degrading purely due to heat issues. And it was audible as well.
I decided to get Mac Studio M4 Max, also all maxed out config and the cooling is so much better that I can run local LLMs like Gemma 3/4, gpt-oss 120b all day long without any heat issues or any audible fan noise. So for my use case it was the right decision. I subsequently added 15'' M5 Max MacBook Pro all maxed out to my collection and even though it is slightly faster on LLM inference (I get 100 tokens/s with Gemma 4 27b model), you just can't run LLMs longer than a few minutes. It starts overheating and gets really loud.
Weird, I’ve run LLM batch sessions for hours on my Max M3 MBP. It doesn’t get very loud, though I’m not getting anything close to 100 tok/s on a 27b model, I use a 35b MoE model just to get 90 tok/s. The fan comes on but thermally it never overheats. I do have it in a vertical closed position, though.
That’s what I have been doing for years, it remains in the house secured while I ssh into it from an old thinkpad. You can get air to pair it with it if you really wanna have that seamless flow, otherwise, ssh works well.
<deleted>
I can't wait for an actual competitor to the M chips from apple. It's frustrating.
What CPU / GPU combination would you recommend? Can use unified host+device memory?
I'd never go x86 again after owning an m1. I'd have replaced an x86 laptop 2 or 3 times by now (2020 m1). My last $3500 dell xps 13z before buying the m1 was absolutely horrible.
Laptops can't do agentic engineering. They get hot as hell and battery drains instantly. I think this will promote a switch to desktops for the next couple of years, until we have new mobile chips.
But laptops can remote into boxes that can run agents.
So a combination of a powerful desktop and a "cheap" laptop might indeed be attractive.
Are you referring specifically to agentic engineering with locally hosted models?
Reading this press release was almost physically painful, because of how much Apple loves to use the phrase "up to". In total it appears 46 times.
up to 46 times, you mean?
10 grand for 256GB memory. Likely double that for 512GB, but won't be available or finalized until October. Thunderbolt 5 is highest bandwidth external IO available at 120Gb/s. 1.2TB/s claimed max internal memory bandwidth.
Not exactly "future proof" for >1T parameter models but good for targeting specific lower-parameter models, or if you can rely on pipeline parallelism and run a cluster.
> Not exactly "future proof"
Computers are never "future proof".
About 20 years ago my dad bought me a $5k computer, it was future proof for about "5 years" before we had to upgrade its internal parts (more memory, new graphics card).
It was future proof but not really because it struggled a lot in its final years.
> Computers are never "future proof".
Upgradeable components however could go a loooong stretch towards that goal. It can't be that hard to follow a common form factor for at least the housing across two or three generations to allow a reuse of everything but the main PCB.
Think it would have same memory bandwidth if the RAM was upgradeable?
Would be nice if someone knowledgeable about electrical engineering and manufacturing processes could lay out some valid reasons for manufacturers to integrate RAM onto the motherboard.
https://news.ycombinator.com/item?id=49041256#49082206
It isn't integrated into the motherboard, it's integrated onto the same package as the CPU/GPU which allows for better signal integrity and higher speeds. You can get somewhat close to the same speeds while modular with tech like LPCAMM2, but there are some pretty difficult challenges to overcome to close the gap completely. As an example, the Framework Laptop 13 Pro CPUs support up to 9600MT/s (same as M5), and Micron sells LPCAMM2 modules that can run at 8533MT/s, but the 13 Pro only officially supports 7467MT/s.
If it was upgradable, then yes, spending more on top of it every year would make it future proof, but that's not the point. It's that spending 10 grand doesn't get you a future proof computer today.
The main PCB is pretty much everything that has value. The rest is a heatsink, case and PSU.
The way the Apple M-series does ram that might be difficult to pull off.
> The way the Apple M-series does ram that might be difficult to pull off.
Well it might be an idea to keep the layout of the mainboard and connectors the same.
That way, instead of having to upgrade the whole machine, all it would need is a new mainboard. Framework for example managed to pull that off, and in mobile at that, where constraints are much worse than for a desktop computer.
> Framework for example managed to pull that off, and in mobile at that, where constraints are much worse than for a desktop computer.
It's not the same thing though. On the M-series, CPU and GPU share a unified memory architecture and ram is much more tightly coupled to get it to go faster. A closer example would be the Framework desktop, actually, where memory is also soldered in for the same reason.
> 10 grand for 256GB memory.
A NVIDIA RTX 6000, 96 GB at 1.7 TB/s, is 13 grand.
This 256 GB at 1.2 TB/s Mac is extremely competitive, it will be sold out everywhere.
The relevant comparison isn't one mac studio to one RTX 6000, it's a 24 channel DDR5 system, which also has ~1.2TB/s of memory bandwidth (or more when Xeon 6 compatible 8800mt/s memory becomes widely available), vastly higher prefill due to more CPU horsepower, orders of magnitude faster networking, can hook into GPU accelerators, can be upgraded etc. A baseline 384GB system from eg Puget is ~30K vs ~12K for the 256GB Mac Studio and you do get value for the money.
So 3x more, plus the cost of a GPU (another 10k?). How is that value for money to get slightly better performance?
It can be more than "slightly", particularly if the model you're interested in (or will be interested in in 6 months) doesn't fit on the mac studio. You also need to account for eg storing 10TB of random checkpoints, load time when experimenting, and so on. When you start actually needing throughput these are all capability gaps in practical use, not just x% benchmark differences.
If you just want to run Qwen 3.8 27B and Deepseek v4 Flash in perpetuity and that's it, there are a lot of solutions that will work and this is a fairly user friendly one.
How's the compute side now, I wonder? Because while the Ultras have impressive memory bandwidth for inference, processing prompts still takes a dog's age on my M3 Ultra. I heard the M5 makes some strides forward in this area, though, and the M7 in particular promises to go a lot further.
M5 is excellent, they’ve finally gotten their own tensor cores.
Good for inference; however if you like to train, data format support and effective performance is limited (M5 Pro). Some hardware features are not exposed or extremely slow.
You’ll be fine for inference, but pales in comparison to what a RTX 6000 Pro can do for compute/matmuls/training.
Except the RTX 6000 will run circles around the Mac studio in just about every way. Memory bandwidth is literally the only spec where Apple is competitive, and while high memory bandwidth is necessary for LLMs to perform well, many people strangely don't understand that memory bandwidth alone is not sufficient.
It better because you’ll need a few of them to run some larger models (I’ll be just as vague citing which models).
3 of those thunderbolt 5 ports, so you can do a fully connected 4 machine cluster topology.
It's unclear to me how bandwidth scales with multiple connections. Many-to-many does not seem ideal. Daisy chaining would be fine for straight pipeline work. There doesn't seem to be an equivalent of a ethernet switch for thunderbolt 5 though.
10 grand for 256GB new Ultra sounds too cheap in today’s crazy DRAM market, it feels too good to be true.
It’d be great when they start releasing these machines with the local models ready for use!
Sorry for being lazy, but is there a rough breakdown like "You get sonnet level for M5 and Opus for M5 pro, etc.", or is it still speculative. Or put simpler, do you get Opus level for the 256GB M5 Max?
For local LLMs with a Mac, rule of thumb is you always want an Ultra (due to memory bandwidth). Even an M1 Ultra is superior to an M6 Pro in this regard.
There are no configurations even close to running something comparable to frontier model variants, they're simply far too large, but something like full precision Qwen 35b or DeepSeek 70b at 50+ t/s is well within available configuration, and potential for plenty of room for large context sizes.
Opus is probably ~2T parameter model, so that would probably not run on these. More like Sonnet.
Sonnet is estimated around 1T, so that is far beyond what's practical as well.
1.2 TB/s bandwidth of M5 Ultra comes from two dies of M5 Max (each 614 GB/s) connected together using 4.4 TB/s inter-die fabric.
For a non-quantized Deepseek V4 flash on an ultra, I would estimate about 1000+ tokens per second prefill and 50+ tokens per second on generation. This is actually quite usable and near parity to cloud.
They mention "adds the GPU Neural Accelerators." which, if exploitable for LLM loads, would probably help the prefill a lot
Yes, and they specifically mention "Up to 10.7x faster LLM prompt processing in LM Studio" which is probably using the neural accelerator for prefill.
It looks like speculation that Apple would raise the base chip’s maximum RAM from 32GB to 48GB was wrong.
Apple also launched the base M6 today with a 32GB RAM limit, suggesting 512GB may remain the maximum for Ultra chips for some time. Since these Ultra chips combine 16 base chips:
32GB × 16 = 512GB
They probably literally don't have enough NAND to go around. 768GB of memory (48GB x 16) is enough for nearly 100 iPhone 17s; that's $800k of iPhones at MSRP, although likely much lower margins than these high-RAM boxes.
They also cancelled availability of the 512 M3 Ultra months ago in many regions, likely just redirecting memory
Mac Studio with M5 Max starts at $2,499 (U.S.) and $2,299 (U.S.) for education. Additional configure-to-order options are available at apple.com/mac-studio. Mac Studio with M5 Ultra starts at $5,499 (U.S.) and $5,099 (U.S.).
I am in Europe, and the Mac Studio M5 Ultra GPU 64 cores with 96GB RAM is up to 6.649,00 €. Ouch.
To put this into a perspective, Google helpfully reminds:
> A fully configured IBM Personal Computer AT (Model 5170) with expanded memory and storage cost around $5,795 to $6,000 at its launch in August 1984, which equals roughly $18,600 to $19,300 in 2026 USD.
It would be interesting to compare the costs of a top of the line machine every decade or so. Costs were steadily decreasing until recently.
John Dvorak said many, many decades ago (80s/90s) that the computer you want will always cost $3000. That statement has been more/less true for some time periods than others, but with some wiggle room I’ve found it to be accurate enough.
Care to guess the approximate price of the MBP I bought earlier this year?
My heavily-upgraded M1 Max came in slightly over that when I got it five years ago. (Still going very, very strong.)
This new Studio? Can't find a config under $5k I'd bother with. But for the MBPs that number still mostly tracks for the average Pro user. (I buy large and run it into the ground so long I mistake the ground for the computer's remains.)
I bought my refurbished M3 Max MBP with 64 GB for $3k a couple of years ago. Before Ethan I never spent more than $2k for a computer.
The “spend” and “want” number might differ, depending on one’s financial state. I know I’ve purchased plenty for less than $3K. But the one I wanted…
They still are, if what you want is roughly the same as the prior generations capability with some uplift (making then number up, but say 20% faster or more ram or whatever).
What is changing is that there genuine demand for more capabilities disproportionate to the cost decrease curve. Fab demand and supply constraints have slowed or even reversed some cost decreases - but that is still getting absorbed by the overall systems costs when you are looking at things like laptops. If you all you want is the last decades demand to browse the web and use office - things are cheaper than ever.
Don't forget that US prices usually do not include the VAT, while EU prices usually do include respective VAT.
A lot of us in America have no sales tax (if that’s what you mean by VAT) and those that do have it at a fraction of EU VAT rates.
> A lot of us in America have no sales tax
Not that many of us, actually. Only Montana, New Hampshire, Oregon, and some parts of Delaware and Alaska have no sales tax. https://commons.wikimedia.org/wiki/File:Sales_tax_by_county....
> I am in Europe, and the Mac Studio M5 Ultra GPU 64 cores with 96GB RAM is up to 6.649,00 €. Ouch.
It would be significantly cheaper to fly to a tariff-free country and buy there.
that's what I'm planning to do.
Where?
The new mac minis and mac studios are going to be in shortage for at least first 6 months from 9.22
> M6 supports up to 32GB of unified memory to multitask across demanding apps and run LLMs on device for secure and private agentic tasks. It also provides up to 170GB/s of unified memory bandwidth — a 10 percent increase over M5 and a 2.5x increase over M1.
Isn't 170GB/s slow for bandwidth?
It is. That's the mac mini. For local LLMs you would want the Mac Studio, which tops out at 1.2TB/s.
M6 and M5 Ultra both have comparable memory bandwidth per GPU core. I think they will perform well.
It's higher bandwidth than any dual-channel DDR5 desktop machine, but Apple never quote the memory latency, so hard to compare otherwise.
Would that be a difficult comparison to do fairly sine one is SoC and the other isn’t?
For max memory bandwidth you need to buy the Ultra versions (M5 Ultra: 1,2TB/s, this gets comparable to real GPUs regarding the memory bandwidth).
It's fast by computer standards and excellent for entry level chip. The Pro/Max/Ultra chips are always faster.
Compared to something like VRAM it's slow.
It's actually quite fast, more bandwidth than NVIDIA RTX 4090 at 1.0 TB/s according to https://www.digitalocean.com/community/tutorials/gpu-memory-... Compute power will still be different, though.
The thread you're replying to is asking about the 170GB/sec memory bandwidth of the base M6, not the faster spec of the Ultra chips.
You would want to get the M5 pro version with 307gb/s if you were interested in running local LLMs.
Kinda. Strix Halo does 256GB/s of memory bandwidth, and is significantly slower than an M5 Max (614GB/s). Feels like intentional market segmentation?
You can get the m5 pro in the Mac mini with 307GB/s at 64gb of memory it’s $2899
Nice for them to add Thread, TB5, and 10G ethernet to Mac Studio. If they allowed first class Linux, this could be a great home server machine.
It's game changer but it's too bad without enough memory
Bizarre there isn’t a 1TB RAM option hidden away for the excessively frivolous or VC funded.
Same reason they cut the big options on the existing models, this way they can sell more devices. The additional cost for the additional 512GB would have to make up for the loss of another sold device otherwise. No idea if there would really be that many people buying this then while on the other hand AI stuff makes people do crazy stuff, so...yeah :)
Considering Apple pricing it actually might.
256GB model is $10k and the 512GB version will probably be double
They seem to be suffering from the supply constraints like everyone else. They phased out the higher capacities on the M3 Ultra Mac Studio a while ago, and if you order a 128GB MBP, say, you're looking at six weeks or more for delivery.
It's more bizarre that Apple got caught with pants down. Focused on CPUs and ignored RAM.
Seems like miscalculation. If they had their own fab for RAM, they could completely corner the market today.
They have no fab for CPUs, they are manufactured by Samsung and TSMC. The bottleneck is in manufacturing RAM not CPUs so there is nothing Apple can do here.
They don’t have their own fab for CPUs either.
They would have had to start building that fab 5-10 years ago. It would have been incredible foresight to do so and it would have looked insane.
I was looking forward and hoping that the Mini and Studio would have 8K at 120Hz. Oh well, maybe the M7s will have that.
512GB unified memory option coming available in October.
The 256GB option is +$4,000 - the overall price for 512GB setup would probably be $20k!
It needs to be a little less than double the overall price for 256GB option, otherwise it's better to get 2 256GB and link them.
you can't link onchip memory.
True, but you can link them up over thunderbolt or Ethernet. If your goal is to run local LLMs, not all weights need to live on the same computer. You can segment the workload by layers and pass the activations along the lower bandwidth interconnect with a small perf penalty. Also you get double the CPU/GPU cores allowing for better multi user/agent performance.
Ah sorry you are right I missread what you meant by it.
M6 (base) in Mac Mini too:
https://www.apple.com/mac-mini/
So weird, launching M6 so silently.
Allegedly, they aren't making a big deal out of the M6 which won't have Pro, Max or Ultra configs at all, instead waiting for the M7: https://9to5mac.com/2026/08/08/apple-m7-chip-heres-why-it-ma...
Interesting, thanks for sharing.
Here's me trying to justify this when I can run frontier models in the cloud for less than the monthly finance charge for this beast.
If you like to experiment with training / finetuning / etc on LLMs, these are actually incredibly ‘cheap’.
1.2TB/s memory bandwidth unlocks a lot with 256GB unified, and agentic AI is pretty good at optimising performance.
For comparison, to get 256GB with NVIDIA, you’re looking at a DIY workstation build (need pcie lanes), and like $70k?
The spark’s ~250gb/s bandwidth doesn’t really count here.
Neither is the right alternative to compare to. You aren’t going to hit 100% utilization (if you are, ignore me, this doesn’t some to you, and write a blogpost for me to read and share).
The comparison should be against renting in the cloud for the duration of your task for training and research or using pay-per-api-call providers for general inference instead of buying your own hardware (and paying the electricity and cooling bills on top), because let’s face it, the models you want to use are probably the same ones available on inference providers (but, yes, some are more trustworthy than others).
Speaking as someone that does ML/AI research, you are essentially paying a huge premium for being able to just run your Python script at any time without setting up a deployment script and harness to run the job remotely, while your hardware sits essentially idle the rest of the time.
The only way to make the math work is if you rent your hardware in the background for inference while you’re not using it in anger, but despite all the startups and promises that has never become as streamlined as mining bitcoins or shitcoins used to be and they don’t pay out as much as they say they would. Renting your hardware for training is another option but doing that is a lot more involved, options are fewer and farther in between, you won’t get as much utilization out of it, and doesn’t let you feasibly abort running tasks at a moment’s notice.
My card (RTX Pro 6000) is always doing something all the time from my queue; like some synthetic dataset generation up next. I still actively use runpods and openrouter for scaled stuff, I was spending a bit and then did the maths, and invested in it.
The maths to me was basically equivalent to prepaying for 242 days of runpod pricing for the same GPU; and I reckon I'd be able to get 6+ years of use out of this card with 96GB.
Plus there's the resell value -- it's actually appreciated by ~50% since I bought it.
Plus I do really enjoy that it's 100% local. I wouldn't feel comfortable giving my agents this much information if inference wasn't 100% local.
I wouldn't get another one, I wouldn't have as much value, but one is definitely paying off for me on the financial side.
To experiment, its still al ot cheaper to prepare everything locally and then just rent a GPU Node on all of these non hyperscalers.
1-2$ / hour.
I'm not regretting my setup at home as it got paid by my company which makes sense here, but paying for electricity is quite high and makes already 0.3$/hour alone.
I would argue, the most interesting use case for running it at home is some personal agent which you want to run 24/7.
Yeah not really lol.
Only reason to buy this if you want to own your compute.
Experimentation and inference are all going to be cheaper on the cloud
512GB unified RAM is going to be really good for running local LLMs.
What is unified ram? RAM and VRAM as one thing?
Big ole pool of very fast ram that can be accessed by the CPU and GPU. Lets you run larger models. AMD does the same thing with Strix Halo. I have a 128gb machine at home, and have had difficulties running 120b models, but 70b and below run pretty well.
Yep, exactly. It's just one shared pool of RAM with zero-copies necessary.
In best case its also on-chip high speed ram.
It seems super reasonably priced to me. It's only twice as expensive as my first Mac which only had 128K of memory.
Depends on your perspective. People think nothing of spending 50-100K on a car that basically gets them to work. But the thing they use for day to day work then gets the evil eye when it costs more than 1K. It's slightly irrational. Not everybody needs a high end mac. But when you do, it sure is nice that you can get one.
I don't actually own a car and my startup is bootstrapped and our salaries are modest. But the one thing we spend on is laptops. I have M4 max pro with 48GB. That thing was on the expensive side (~4.5Kish). But it delivers a lot of value and I spend most hours I'm awake using it. I like fast builds. I like that I can try out open source AI models. And I like just having the option to run those.
We actually lease them and mine costs something like 105 euro/month. Including Apple Care. I don't need a Mac Studio but I could see some roles where that would not be a crazy expense. Even the tricked out version that basically only costs the same as a very modest car.
Base models are okay-ish.
But +4000$ for an additional 128GB of ram is simply milking the customers, as they know they will have many of them.
That came with a monitor and floppy drive.
and this one comes without a floppy drive.
Very much happy with my machine.Got a base m4 max studio in December. 2350eur what a deal
I only with I got more than 96GB at the time.
Boy, oh boy Apple is the new shovel seller during AI gold rush.
Most people at Apple have already realized that their processors are already too powerful for regular users - heck, as a developer my M2 Pro with 32 GB RAM is more than enough for me.
Regular users don’t care about local AI either. So, they will probably extract as much money as possible during AI gold rush, but then we will most likely see Apple
a. Making their software worse (god forbid, forced updates)
b. Making their hardware impossible to repair (as they almost accomplished this already) and easier to break.
> Boy, oh boy Apple is the new shovel seller during AI gold rush.
They're eating nvidia's lunch.
I dunno, Apple's been throwing many bones to their customers who use their Macs for AI stuff; like Apple working with, and signing TinyGPU's NVIDIA eGPU drivers.
Plus introducing features like RDMA over thunderbolt, which is critical for distributed inference/training/etc. On the software side, Apple is investing heaps.
It's still ridiculous they don't support expandable NVMes, but the memory being soldered makes sense, you need it for 1.2TB/s bandwidth.
They are still selling high-margin hardware. Apple loves selling high-margin hardware.
Wow, didn't know this exists. So instead of ditching my M1 Pro Macbook for a newer one with more RAM and power I could do this instead (if reasonable regarding pricing).
EDIT: AMD too, it's not limited to Nvidia, nice.
No looking to cheat, which is better: the MAX or ULTRA?
Depends if it's Pro Ultra Max, or Plus Max Ultra.
I wish they were offering 1TB of Unified Memory for the M5 Ultra. I already have an M5 Max MBP w/ 128GB of RAM for running local models, and while there's a /few/ models that I can run in 512GB that I can't run in 128GB that are interesting, where things really shift is at 1TB of memory which allows you run >1T parameter models w/ 4 bit quants reliably. 512GB is just on the edge of "enough", which is maybe the point of maximum frustration considering current memory prices.
Personally, I can't justify dropping the dosh for a 512GB M5 Ultra, but I would be able to justify it to myself if I could get 1TB of memory, because it'd guarantee the flexibility with local models I currently am missing. Seems a huge miss to not offer this... for a price.
1tb would likely be ~$20k - given the current >$10k price tag of 256gb. Would you still be considering it at that price?
I'd consider a 1tb machine at 20k, but I'm not going to pick up a 256gb one at all. 1TB fits a frontier-ish model in memory without massive quantization, which is a very interesting capability for a non-rack piece of compute.
But why...?
At that point just rent proper GPUs in the cloud, you'd have way more power and pay only what you use for.
More likely double that, even. I think you'd still see many buyers there. You can spend like $16k alone on a RTX 6000 PRO with a mere 96GB of VRAM now..
I would probably spend up to $30k if I could get 1TB of Unified Memory, because it would allow me a guarantee to run pretty much any local model I want, including >1T parameter models with reasonable quants. I wouldn't be surprised if 512GB is close to $20k when it becomes orderable in October. The justification is less about absolute price and more about price to what it enables. 512GB really doesn't enable much over 128GB for me, but 1TB would massively change things.
I have a bit of paranoia/anxiety about AI, but it's not what most people are concerned with. I understand the limits of these tools very well, and still find them extremely useful. What concerns me is that it's going to become difficult to impossible in the future to run local models which have near-SOTA capabilities in a way in which you can exercise full control of the model. I see the writing on the wall, and its more than worth it for me to invest early to ensure my own capabilities. I am very much not a fan of our "you'll own nothing and be happy" directionality for the world, and I am (at least currently) privileged to have the means to slow that decline for my own self.
chaining an option?
RDMA is buggy and Thunderbolt only delivers 1/10th the throughput of native connectivity. 1TB of Unified Memory w/ 1.2TB/s of bandwidth with marginally ~$30k cost is a different story than 1TB of sorta Unified Memory w/ an effective 120GB/s of bandwidth with a marginally ~$40k cost + all the RDMA bugs.
You need latency for token parallelism, not bandwidth. Hence actual RDMA that bypasses the software TCP stack (ROCe or whatever).
If I were to get one of these, realistically what is the most advanced AI model I could run locally?
GLM 5.3 in ~q4 quantiziation with 4bits
"512GB memory option for M5 Ultra coming late October"
anyone know if this is "pre-order" is coming late October, or "will be available to ship in late October, thus pre-order will be available earlier than that"
Am I the only one that now finds press releases like this similar to "AI Slop"
I know there's tons of marketing language, buzz words and attempts at convincing me of some agenda that isn't super clear without lots of effort in "validating" the slop. I guess its not bad "slop" though if a human put in effort in editing it (imo >50% human curating = not really bad ai slop)
Though I still would prefer I could just get the prompt. What human thoughts, direction and "prompt" went into writing this article? in the same way as we ask for the prompt for AI generated outputs, I would prefer it for human generated output too. For writing at the least. I could have saved time, got the purity of the argument, and got more clear information. I wonder if we can get a future where humans just express their intent with each other and stop trying to hide our agenda; I want a world we can trust each other greater and interpret and act on our goals without the noise of trying to impress or market to each other & the additional words that go into that.
The article's headline contains an em dash. I wonder if it's AI-generated
My sarcasm detector just made a small cloud of smoke. What does that mean, a buffer over or underflow??
To me it seems like a cheeky way to preface the AI part of the title.
$5500 for 96GB of RAM? insanely expensive (the MacOs is really bad compared to Windows or Linux).