5 years of (200/month) tokens at that price, meanwhile an rtx 5090 pc is about half that… hmm
but i wonder how much these token costs are sustainable or not, it may be in the long term cheaper to have your own hardware if token costs go up (and hopefully hardware gets cheaper again)
The token price isn't the only reason to run a model locally though. You can do additional training to specialize or remove censorship that may be a no-no per TOS with cloud GPUs.
They do MoE. They benchmarked GLM 5.3-flash (320B / 18B), and Qwen 3.8-flash-next (125B / 6B). The dense Qwen is only focused (I assume) because it's about the only thing that fits on a 5090, that they can compare the two heads on.
A dense model (up to the amount of memory available) actually does make the most sense on unified memory architectures. But when you hit the limit of what you can hold in memory, you reach the limitation of the platform.
Whereas a hybrid architecture with distinct DRAM and VRAM with sparse MoE, you can leverage two different bit rates depending on the actual need for constant access to common layers versus sparse access to infrequent layers and arbitrage the difference in cost for each of those in distinct classes of hardware.
This 1000%. Data centres don't equate to medium sized labs and businesses. A stack of Macs is up and running without digging trenches, an electrician on staff and a department of PhDs to justify the spend.
It's likely that a stack of Macs will draw more power for slower prefill/decode than equivalently priced Nvidia GPUs. If power efficient inference is the goal, Macs are a non-starter.
A) the macos value add is enormous if you have any investment in the ecosystem, B) for me at least a GPU is completely useless for anything but being a token generator.
A 5090 has a 1.79TB/s memory bandwidth. Qwen 3.8 27B NVFP4 is 22GB. You cannot generate tokens faster than the weights can traverse the GPU memory, so that makes max generation speed without MTP to be 81T/s. Say MTP is giving you 0.5 acceptance rate (very good), that is 1.5 * 81 is 121T/s. Even with a perfect acceptance rate you would only get 162T/s.
Off the top of my head, I'm guessing we're missing sparse attention. But I'll run your challenge through and see where the gaps are. I promise I'm telling the truth :)
Have a 5090, and yes it's very fast. But it's like the worst ADHD team member and requires constant supervision and review from larger models. It's context size on-card is good for super, suuuuuper shallow precision work.
The gb10/spark on top of it, that thing can refactor enormous monorepo architecture. The time it takes the 5090 to compact, reiterate and execute a plan is often the same time as the gb10.
When I say that Apple is astroturfing HN, this is what I mean. Anyone can easily verify that even a 3090 will blow any Mac out of the water in terms of Tok/sec. Somehow its ok to just post outright lies when it comes to Apple product comparison.
Thank you for this. I wish Apple focused their silicon design on improving the TTFT metrics but coming from an M3 Pro, it still looks laggard compared to Nvidia's TensorCores in the 5090.
Maybe Apple is an acquisition away from changing that balance.
The rumor on Apple's processor roadmap is that they're skipping other M6 variations (all previous generations had Pro and Max, a few had Ultra) in order to focus on the M7 generation for AI reasons. What exactly the M7 improvements are who knows.
I think that comes down to TSMC. Nvidia apparently booked out the whole A18 or 16 node. Apple is on 2nm right now and M7 will jump right to A14. According to my quick AI research anyway.
The selling point of the M5 Ultra Mac Studio is that you can run much larger models that the 5090 can't without swapping. NVidia aggressively segments the market on VRAM for this reason. That's why a 5090 has an MSRP of ~$2k (but good luck getting one for less than $4k) while a 6000 Pro, which is basically a 5090 with 96GB of RAM has now soared beyond $15k where 3-6 months ago it was more like $10-11k. A 6000 Pro has the same memory bandwidth but slightly more CUDA units (IIRC ~24k vs ~21k).
This advantage won't be apparent with a 27B model. The 256GB MS can probably run the newer Flash models locally, something you can't do on a 5090.
I don't think we'll get a successor to the 5090 until late 2028, maybe even 2029. I'm basing this on the launch date of the 5000 series and that we haven't got a midcycle refresh yet. Rumor has it the chips are ready but the 3GB RAM modules are 3-4x the price of the 2GB modules used on the current cards.
Apple should see a Mac Studio major update in 2028. That might even force NVidia's hand. But it's really impossible to say what the state of the market will be 2-3 years from now. It may have completely crashed. I suspect not however.
The interesting thing will be when the bandwidth demands start forcing HBM memory onto these home/enthusiast solutions.
But what about builds that combine 8 of the 5090 with infiniband between boxes? Wouldn't that be comparable to the mac in terms of price and potentially beat it by a lot in terms of performance for the large MoE? I understand the space/heat/noise considerations, but price wise it may still not make as much sense as people think. (Agreed that it is hard to get the NVIDIA hardware and the 6000 pro are priced less competitively).
While that sounds super awesome,
How many people are actually going to build and maintain that vs a box you can grab at the mall that fits in a lunchbox?
I can't speak to Infiniband pricing for something like that. It seems like the cheap option is 56/100Gbps with used Enterprise equipment. You'd need 8 HCAs, DAC cabling and a switch but even then you're into thousands of dollars. If you want 200Gbps+ it gets into the tens of thousands (AFAICT).
Each PC is probably going to cost ~$6k and you're talking about 8000W of electricity draw. That's going to consume multiple 20A circuits even at 240V. And the electricity ain't free either. A Mac Studio seems to draw ~500W max.
Oh and the Mac Studio has an upgrade route to run 1T+ models too by chaining them together with TB5 chaining. OSX supports RDMA this way. That's comparable bandwidth to the 100Gbps Infiniband option.
So you're talking about $50-60k of hardware and more power draw and more heat for something that will I'm sure beat the MS M5U option but at huge cost. Also, at that kind of price point, I'm likely to get a workstation PC and put 2 (or possibly 3) 6000 Pros in it.
Certainly not forgetting wattage. A 5090 is 575W. The M5 Ultra Studio is 480W.
nvidia-smi -pl 450 for like a 4% reduction in throughput. I tend to set it around 350W because it's a comfortable temperature blowing on my legs under the desk without warming my office in the summer.
I put together this system two years ago, so it's a little out of date, but it only cost $3000 for the same performance and capability as an Ultra. I don't think I would spend $7000 to save 100W, though.
Now try running that Qwen 3.8 Next model on the 5090 and tell me what TPS you get (hint: it's near 0 since it doesnt fit the 32GB VRAM on 5090 vs the 256 in OPs M5).
Surprisingly, the Reddit crowd are reporting 50–60 tokens/s (for the 32 GiB 5090 + 128 GiB RAM)—on par with the M5 Ultra benchmarks, despite both the PCIe bottleneck and much smaller DDR5 bandwidth,
I assume those are non-batched. I think the M series GPU can do 4X to 8X depending on model quant, which means if you can batch queries you'll get almost 4X to 8X performance.
On my m5 max 27b model does 75tps on 256k ctx and starts at 80 on the 8k ctx when you add https://huggingface.co/collections/z-lab/dflash-2 to it. So yeah base might be 30tps (I used iq4) but mtp or dflash help a lot and should be used when checking what is useful and what is not for running models as it is not fare to judge without them.
On Apple website it says 512GB memory option is available in October. I guess bumping to that one would cost additional 4-6k US$. So an Ultra with 2TB storage would be north of 15k US$.
That’s like 12 years worth of OpenAI Pro subscriptions
I think we all expect the heavy subsidized subscriptions to end or significantly increase in price at some point, but it could be years from now and I'd rather spend a similar figure on an hypotetical Mac Studio M8 Ultra, or whatever more advanced competitor that will have likley appeared by that time.
A more apples-to-apples comparison would be with API cost in OpenRouter at the same tok/s rate for the same models that you can run locally, maybe.
HN always has these completely contrived counterarguments. What is actually going to realistically happen that will prevent use of an LLM provider? Did you think that the OP literally meant the 12 years or maybe it was just to show how expensive using a Mac Mini as an alternative is?
Yes, it feels like that. Whereas frontier labs are pushing the frontier of human knowledge, selflessly working towards pulling humanity from dark ages. Ignorant peasants trying to burn the modern civilization down. Don't they know data centers and fiber lines are lifeline of modern economy?
how expensive using a Mac Mini as an alternative is?
I think it goes without saying. And it is eminently evident over last couple of decades that from compute to storage to meals 3rd part providers have saved billions upon billions of dollars to enterprises and individuals alike by providing these essential services.
Yeah, anyone who thinks local AI is going to save them money is likely to be disappointed, at least if they want to run models that are even remotely capable.
Plenty of other reasons to get excited about local AI, but I don't think cost is one of them.
Maybe you are using a local model to go after some Millennium Prize problem and you don't want OpenAI to take your work and use it to win the prize for themselves? $15k might be a bargain.
And, yes, I know a current local model wasn't going to solve the Navier-Stokes problem, but I'm just using it as an example where privacy might be valuable.
Despite being on a site called Hacker News, we seem to often overlook the simple aspect of wanting local AI hardware to hack (not necessarily in the cybersecurity sense) with. I got my local AI hardware because it's an enjoyable hobby for me.
Hard to guess, it can go either way. If you will need to be in a syndicate to use non-sterilized models, that mac makes sense. But if there is mandatory registration of personal cyberarms, you risk going to mines once they check you purchases. You could try to play normie and pretend you simply wanted to show off, by keeping your actual work on external disk, but that leaves traces on system. Counting on someone in the Gap renting you gray iron works as long as you can swap credits. Still, this gear is tiny. Put it in your e-car, with uplink, and leave it at uncle's farm. Discreet.
Just commenting here because you're discussing hardware: I thought the test results from the SSD published in this article [1] were pretty interesting. Maybe that's old news though.
512 option isnt worth it imo, you get severe slowdowns when weights are that large. 256 is the sweet spot, you can run large open weight models at decent speeds for full private inference.
"It also happens to be a Mac, with an operating system that looks nice and doesn’t suck"
Yes Apple has some of the best hardware out there, albeit overpriced. But the software is such a hindrance and I can't take anyone that states otherwise seriously. If only it had proper Linux support (and the Asahi people do an amazing job but you can reverse-engineer only so many stuff with limited funding, and then you have to do it again for new models). MacOS is good if you just want to have a standard experience, which to be fair is most people. It's good for just setting up an LLM server I guess since the hardware is a perfect fit. I wouldn't touch it otherwise.
This is great as a first look, but the author is not a developer, so we don't yet know whether a dev can be as productive with local models on M5 Mac Studio compared to a 20x subscription plan.
I'm also curious about any new low hanging optimization opportunities in the kernels for this new hardware.
It's already clear to me that M5 Mac Studio is more cost-effective than anything you can run on open router, assuming decent utilization.
The M5 Mac Studio will be the most cost effective way to run uncensored cyber capable open agents.
An exciting tipping point will be if programmers can get an Astra-Ultra like experience all week with this hardware. That would be a real sense where this hardware exceeds the value of even 20x cloud subscriptions.
This is great as a first look, but the author is not a developer, so we don't yet know whether a dev can be as productive with local models on M5 Mac Studio compared to a 20x subscription plan.
Local models are definitely not as productive as SOTA, sadly it's not close yet. I do think someday they will be "good enough" to use, but they aren't today. Even the SOTA models barely code well, with Opus 4.5 being the first, good coding model.
That being said, I think it's absolutely imperative that we keep pushing local model performance. We need to continue to advance technology there and ensure that the model labs don't do regulatory capture in the name of "safety" (or anything else).
With the latest codex (weekly quota burn) fiasco I tried open weight alternatives for the first time. And tyeah... open weight models cant compete with likes of astra yet. But, my hope is that by the time I get my Mac studio at end of november an open weight models would have closed the gap (which i think is realistic at the speed of progress). Now its true a better gpt version will also be available then but it also seems the gap is shrinking with time so theres that.
Local models can be widely used as productive assets. Yes the infrastructure of SOTA API models is engineered specifically for you to be that utility, but the blanket statement that local isn't up to par is intensely short sighted. Billions of tokens per month on local pays for the hardware when compared to sota costs per month.
I believe they can currently be used productively for non-coding tasks (classification, light summary)... but they definitely are not even close to SOTA when it comes to software development.
Defining productivity is a use-case scenario, and a wildly generalized assumption for most people in this argument. Local infrastructure doesn't need to be sota for absolutely every single need for a dev lab, but it absolutely can be delivered with non-api frontier class models.
Just to be clear, I'm specifically talking about coding. I think local models can help with productivity today, just not coding.
I'm also a huge fan of local models and think it's absolutely imperative that they continue to advance so we can move off of the Anthropic/OpenAI hosted models. It's important to accurately asses where we are in that journey though.
Astra-Ultra? Even the largest open model to date (Kimi K3) is nowhere close to Astra level, and it will be quite slow even on the highest-spec M5 Ultra, with achievable speeds of about 0.5 tok/s at most due to having to stream weights from SSD (~13 GB/s on the highest storage capacity M5 Max machines so far). This is OK for doing simple Q&A in the background but it's far from a genuine coding experience. You'd have to test batching of multiple thinking streams in order to try and raise overall tok/s via layer-wise reuse of the streamed weights (and this is where the "Ultra" part sort of becomes relevant; Kimi series models have good support for agent swarms) but this would decrease single-session performance even further. It would only be usable for background jobs, though the hardware would then have a chance of paying for itself if it was fully used on a 24/7 basis.
While I know it's not apples to apples, the target comparison right now is 2x DGX Sparks. Similar price, 256gb. The conversation has focused on memory bandwidth vs. compute in agentic loops, so for most people the raw numbers will mean less than the "time per task" in coding benchmarks.
This is a great article and bodes well for the M5, but we should expect more like this comparing to other platforms before we truly understand where it fits.
Speed vs task-completion is a new conversation and a great point. Whereas the cost to compute doesn't exist in a vacuum, making mistakes costs less, is easier to maintain with granularity and a whole host of other factors when you own the lab.
It's not the M5 Ultra itself, but the M7s or M9s that will do the damage.
99% of people will use whatever AI is free. The sophisticated, heavy users that are willing and able to pay a lot of money the ones that will be interested in controlling their inference bills.
Today, the sweet spot where an M5 Ultra makes sense is tiny. But we might expect that to grow a lot.
Even if you could get a frontier model, you would not be able to run it on any Mac. So speculating on what M7 or M9 will achieve in 5 years (if we even still exist) seems pointless.
You will. Some of us are going to be already ground into dust that the microwave parts are made out of. Others will be locked into our communism cubes with our daily allotment of entertainment and sustinece. let out into the sunlight for only 30 minutes per day.
Ordered one for OpenFOAM. Excited for it. Will be nice to not have my laptop running CFD 24 hours a day, but my M1 Max is currently my fastest machine… I’m expecting about 3.5x from the M5 Ultra.
I think Apple is really sleeping on making this run a Linux server. These things are very capable and draw very little wattage when idle. It would make an excellent homelab device, but MacOS currently holds it back in this regard.
My mac is 5 years old. I don't think I can comfortably buy a new one right now. It has a 16GB unified RAM. Honestly that would be enough for so many local models that I want to use but can't use. Because RAM usage (even with literally every single user installed app quit/stopped) the RAM usage is very high that I can barely safely get 6-7 GB (I am supposed to get ~10 GB, but it goes up and down real fast!). That's a shame. If only I could install an alternative OS that uses very little amount of RAM :-)
When people benchmark MLX related quant models, they really need to publish numbers on benchmarks. You cannot take this as it is what you get of the original models. MLX uses pretty simple quantization methods so at lower bits without QAT, it is just not as good quality as llama.cpp ones.
At this point I think I will get the DGX gb300 workstation though I will wait a bit more for the cold season. It is double the price but at least is the real thing
I think we'll eventually get to the point where folks will have a local AI agent but I think people need to temper their expectations to a degree. You aren't going to have data center level tok/s from a box sitting under your desk and you don't need instantaneous responses for many workloads. Having a local agent that can execute tasks over a couple days with your supervision that might otherwise take you weeks is perfectly acceptable.
However I also think that Agentic AI is very much not an out-of-the-box solution, local or otherwise, and it takes a high level of technical knowledge to create an effective AI agent. And there's a problem now where most orchestration is fixed on what models are used for what tasks with no ability to weight constraints like cost, speed, and security.
$12,299
5 years of (200/month) tokens at that price, meanwhile an rtx 5090 pc is about half that… hmm
but i wonder how much these token costs are sustainable or not, it may be in the long term cheaper to have your own hardware if token costs go up (and hopefully hardware gets cheaper again)
The token price isn't the only reason to run a model locally though. You can do additional training to specialize or remove censorship that may be a no-no per TOS with cloud GPUs.
The numbers I was most interested in are tucked away in a chart towards the bottom - the speed comparison of the Mac Studios v.s. a RTX 5090:
A whole bunch more comparison numbers in this section: https://www.macstories.net/stories/m5-ultra-mac-studio-revie...
Those are some incredible graphs, that leap in prompt processing going from M3 to M5.
Also: ~30 token/s on GLM 5.3-flash, locally. (That's roughly Opus 4.8-tier. I think).
/meta Here's a CSS filter that stops those nuisance chart animations,
A dense 27B doesn't really make sense for the Mac. A MoE makes way more sense when you have modest bandwidth but lots of memory.
They do MoE. They benchmarked GLM 5.3-flash (320B / 18B), and Qwen 3.8-flash-next (125B / 6B). The dense Qwen is only focused (I assume) because it's about the only thing that fits on a 5090, that they can compare the two heads on.
A dense model (up to the amount of memory available) actually does make the most sense on unified memory architectures. But when you hit the limit of what you can hold in memory, you reach the limitation of the platform.
Whereas a hybrid architecture with distinct DRAM and VRAM with sparse MoE, you can leverage two different bit rates depending on the actual need for constant access to common layers versus sparse access to infrequent layers and arbitrage the difference in cost for each of those in distinct classes of hardware.
Those RTX 5090 numbers are bad. You can get over 200 tps with ninfer using NVFP4 and MTP.
can confirm.
I dont' know why people spend huge money on these and Spark. The 5090 is running qwen 3.8 at 200+ tps!! That's 1-2 orders of magnitude faster.
same reason they spend huge amounts of money on rolexes when seikos work better (the tech crowd isn't immune from vanity).
If you seriously think apple products are nothing but a status item, you're deluding yourself and probably have been for decades.
This 1000%. Data centres don't equate to medium sized labs and businesses. A stack of Macs is up and running without digging trenches, an electrician on staff and a department of PhDs to justify the spend.
It's likely that a stack of Macs will draw more power for slower prefill/decode than equivalently priced Nvidia GPUs. If power efficient inference is the goal, Macs are a non-starter.
So if it isn't a comparative ability, now it's a power cost issue? This reads like goal post moving.
If you seriously think apple cares about anything other than cell phones, you're deluding yourself and probably have been for decades.
They've been selling phones for less than 20 years at this point? Though I suppose 1.9 is not equal to 1, so it gets the plural.
People don't buy Sparks and M5 Ultras to run a 27B model - you buy it to run an MoE model like Qwen Next which this M5 excelled at.
A) the macos value add is enormous if you have any investment in the ecosystem, B) for me at least a GPU is completely useless for anything but being a token generator.
No thanks to the "macos value add" that forces you to use Metal while Valve customers frolick in Protonland.
A 5090 has a 1.79TB/s memory bandwidth. Qwen 3.8 27B NVFP4 is 22GB. You cannot generate tokens faster than the weights can traverse the GPU memory, so that makes max generation speed without MTP to be 81T/s. Say MTP is giving you 0.5 acceptance rate (very good), that is 1.5 * 81 is 121T/s. Even with a perfect acceptance rate you would only get 162T/s.
Off the top of my head, I'm guessing we're missing sparse attention. But I'll run your challenge through and see where the gaps are. I promise I'm telling the truth :)
Have a 5090, and yes it's very fast. But it's like the worst ADHD team member and requires constant supervision and review from larger models. It's context size on-card is good for super, suuuuuper shallow precision work. The gb10/spark on top of it, that thing can refactor enormous monorepo architecture. The time it takes the 5090 to compact, reiterate and execute a plan is often the same time as the gb10.
Both are probably single-token decode performance, which is reasonable to show. Otherwise agree RTX 5090 should shinebetter with NVFP4.
When I say that Apple is astroturfing HN, this is what I mean. Anyone can easily verify that even a 3090 will blow any Mac out of the water in terms of Tok/sec. Somehow its ok to just post outright lies when it comes to Apple product comparison.
Thank you for this. I wish Apple focused their silicon design on improving the TTFT metrics but coming from an M3 Pro, it still looks laggard compared to Nvidia's TensorCores in the 5090.
Maybe Apple is an acquisition away from changing that balance.
The rumor on Apple's processor roadmap is that they're skipping other M6 variations (all previous generations had Pro and Max, a few had Ultra) in order to focus on the M7 generation for AI reasons. What exactly the M7 improvements are who knows.
https://www.macrumors.com/2026/06/25/2027-macs-m7-chips/
I think that comes down to TSMC. Nvidia apparently booked out the whole A18 or 16 node. Apple is on 2nm right now and M7 will jump right to A14. According to my quick AI research anyway.
That sounds a lot like AI fantasy slop.
Apple just shifted to N2. They’re not going to be doing another major shift right away.
And TSMCs own roadmap would put your hallucination years away at best for a a product that follows a roughly annual cadence https://www.tomshardware.com/tech-industry/semiconductors/ts...
The selling point of the M5 Ultra Mac Studio is that you can run much larger models that the 5090 can't without swapping. NVidia aggressively segments the market on VRAM for this reason. That's why a 5090 has an MSRP of ~$2k (but good luck getting one for less than $4k) while a 6000 Pro, which is basically a 5090 with 96GB of RAM has now soared beyond $15k where 3-6 months ago it was more like $10-11k. A 6000 Pro has the same memory bandwidth but slightly more CUDA units (IIRC ~24k vs ~21k).
This advantage won't be apparent with a 27B model. The 256GB MS can probably run the newer Flash models locally, something you can't do on a 5090.
I don't think we'll get a successor to the 5090 until late 2028, maybe even 2029. I'm basing this on the launch date of the 5000 series and that we haven't got a midcycle refresh yet. Rumor has it the chips are ready but the 3GB RAM modules are 3-4x the price of the 2GB modules used on the current cards.
Apple should see a Mac Studio major update in 2028. That might even force NVidia's hand. But it's really impossible to say what the state of the market will be 2-3 years from now. It may have completely crashed. I suspect not however.
The interesting thing will be when the bandwidth demands start forcing HBM memory onto these home/enthusiast solutions.
But what about builds that combine 8 of the 5090 with infiniband between boxes? Wouldn't that be comparable to the mac in terms of price and potentially beat it by a lot in terms of performance for the large MoE? I understand the space/heat/noise considerations, but price wise it may still not make as much sense as people think. (Agreed that it is hard to get the NVIDIA hardware and the 6000 pro are priced less competitively).
While that sounds super awesome, How many people are actually going to build and maintain that vs a box you can grab at the mall that fits in a lunchbox?
No, $40K is not comparable to $10K.
Sounds like nice utility bill in the making.
I can't speak to Infiniband pricing for something like that. It seems like the cheap option is 56/100Gbps with used Enterprise equipment. You'd need 8 HCAs, DAC cabling and a switch but even then you're into thousands of dollars. If you want 200Gbps+ it gets into the tens of thousands (AFAICT).
Each PC is probably going to cost ~$6k and you're talking about 8000W of electricity draw. That's going to consume multiple 20A circuits even at 240V. And the electricity ain't free either. A Mac Studio seems to draw ~500W max.
Oh and the Mac Studio has an upgrade route to run 1T+ models too by chaining them together with TB5 chaining. OSX supports RDMA this way. That's comparable bandwidth to the 100Gbps Infiniband option.
So you're talking about $50-60k of hardware and more power draw and more heat for something that will I'm sure beat the MS M5U option but at huge cost. Also, at that kind of price point, I'm likely to get a workstation PC and put 2 (or possibly 3) 6000 Pros in it.
Not forgetting of course that an RTX5090 is what 600W+ ? And the Mac is probably half that at most ?
sure. so is 2x power worth 10x perf? I think it is in most cases.
Certainly not forgetting wattage. A 5090 is 575W. The M5 Ultra Studio is 480W.
nvidia-smi -pl 450 for like a 4% reduction in throughput. I tend to set it around 350W because it's a comfortable temperature blowing on my legs under the desk without warming my office in the summer.
I put together this system two years ago, so it's a little out of date, but it only cost $3000 for the same performance and capability as an Ultra. I don't think I would spend $7000 to save 100W, though.
That's a dense model. Of course it will do worse.
Now try running that Qwen 3.8 Next model on the 5090 and tell me what TPS you get (hint: it's near 0 since it doesnt fit the 32GB VRAM on 5090 vs the 256 in OPs M5).
Surprisingly, the Reddit crowd are reporting 50–60 tokens/s (for the 32 GiB 5090 + 128 GiB RAM)—on par with the M5 Ultra benchmarks, despite both the PCIe bottleneck and much smaller DDR5 bandwidth,
https://old.reddit.com/r/LocalLLaMA/comments/1wl06np/qwen38f...
(Note it's a sparse MoE with only 6B active).
Good to know thanks.
That's with CPU offload to a DDR5 6000 RAM though which is around $3-4k at least.
I assume those are non-batched. I think the M series GPU can do 4X to 8X depending on model quant, which means if you can batch queries you'll get almost 4X to 8X performance.
On my m5 max 27b model does 75tps on 256k ctx and starts at 80 on the 8k ctx when you add https://huggingface.co/collections/z-lab/dflash-2 to it. So yeah base might be 30tps (I used iq4) but mtp or dflash help a lot and should be used when checking what is useful and what is not for running models as it is not fare to judge without them.
On Apple website it says 512GB memory option is available in October. I guess bumping to that one would cost additional 4-6k US$. So an Ultra with 2TB storage would be north of 15k US$.
That’s like 12 years worth of OpenAI Pro subscriptions
Agreed.
Specially since one can pay half right now to OpenAI and sign a 12 year iron clad contract for uninterrupted service delivery of OpenAI Pro.
I hope that was sarcasm.
"iron clad" :)
Censorship included.
I think we all expect the heavy subsidized subscriptions to end or significantly increase in price at some point, but it could be years from now and I'd rather spend a similar figure on an hypotetical Mac Studio M8 Ultra, or whatever more advanced competitor that will have likley appeared by that time.
A more apples-to-apples comparison would be with API cost in OpenRouter at the same tok/s rate for the same models that you can run locally, maybe.
Don't you mean an Apple to NVidea comparison?
HN always has these completely contrived counterarguments. What is actually going to realistically happen that will prevent use of an LLM provider? Did you think that the OP literally meant the 12 years or maybe it was just to show how expensive using a Mac Mini as an alternative is?
Mass revolts of the peasantry burning down data centers and cutting fiber lines.
Yes, it feels like that. Whereas frontier labs are pushing the frontier of human knowledge, selflessly working towards pulling humanity from dark ages. Ignorant peasants trying to burn the modern civilization down. Don't they know data centers and fiber lines are lifeline of modern economy?
I think it goes without saying. And it is eminently evident over last couple of decades that from compute to storage to meals 3rd part providers have saved billions upon billions of dollars to enterprises and individuals alike by providing these essential services.
Yeah, anyone who thinks local AI is going to save them money is likely to be disappointed, at least if they want to run models that are even remotely capable.
Plenty of other reasons to get excited about local AI, but I don't think cost is one of them.
Maybe you are using a local model to go after some Millennium Prize problem and you don't want OpenAI to take your work and use it to win the prize for themselves? $15k might be a bargain.
And, yes, I know a current local model wasn't going to solve the Navier-Stokes problem, but I'm just using it as an example where privacy might be valuable.
Agreed, plenty of other reasons to get excited about local AI.
Despite being on a site called Hacker News, we seem to often overlook the simple aspect of wanting local AI hardware to hack (not necessarily in the cybersecurity sense) with. I got my local AI hardware because it's an enjoyable hobby for me.
Hard to guess, it can go either way. If you will need to be in a syndicate to use non-sterilized models, that mac makes sense. But if there is mandatory registration of personal cyberarms, you risk going to mines once they check you purchases. You could try to play normie and pretend you simply wanted to show off, by keeping your actual work on external disk, but that leaves traces on system. Counting on someone in the Gap renting you gray iron works as long as you can swap credits. Still, this gear is tiny. Put it in your e-car, with uplink, and leave it at uncle's farm. Discreet.
I gotta have some of what you had :)
I thought it was a fun bit of cyberpunk fiction. Those who downvoted him seem to have taken it at face value?
I appreciate the reference to RUSH: Red Barchetta in the final line.
Just commenting here because you're discussing hardware: I thought the test results from the SSD published in this article [1] were pretty interesting. Maybe that's old news though.
[1] https://www.macworld.com/article/3238319/mac-studio-m5-max-r...
512 option isnt worth it imo, you get severe slowdowns when weights are that large. 256 is the sweet spot, you can run large open weight models at decent speeds for full private inference.
The model being tested is 18k as configured.
I didn't expect this to make the 5090 to look like a good deal.
5090 has 32GB VRAM.
It'd be silly to buy the 18k model to run a tiny model like Qwen 27B. You use models like GLM Flash and Qwen Next which won't fit on a single 5090.
Is it that silly? You could run multiple 27B models in parallel.
You actually don't need more RAM to batch multiple inference tasks of the same model.
(Each task needs its own context, but the (e.g.) 27B of constant parameters isn't duplicated).
Yes Apple has some of the best hardware out there, albeit overpriced. But the software is such a hindrance and I can't take anyone that states otherwise seriously. If only it had proper Linux support (and the Asahi people do an amazing job but you can reverse-engineer only so many stuff with limited funding, and then you have to do it again for new models). MacOS is good if you just want to have a standard experience, which to be fair is most people. It's good for just setting up an LLM server I guess since the hardware is a perfect fit. I wouldn't touch it otherwise.
What exactly is missing from macOS that makes you feel the need for Linux?
I get it on Windows systems, at least when someone wants to use Linux-type tooling. But macOS already supports pretty much all of that natively?
For starters, the source code.
This is great as a first look, but the author is not a developer, so we don't yet know whether a dev can be as productive with local models on M5 Mac Studio compared to a 20x subscription plan.
I'm also curious about any new low hanging optimization opportunities in the kernels for this new hardware.
It's already clear to me that M5 Mac Studio is more cost-effective than anything you can run on open router, assuming decent utilization.
The M5 Mac Studio will be the most cost effective way to run uncensored cyber capable open agents.
An exciting tipping point will be if programmers can get an Astra-Ultra like experience all week with this hardware. That would be a real sense where this hardware exceeds the value of even 20x cloud subscriptions.
Local models are definitely not as productive as SOTA, sadly it's not close yet. I do think someday they will be "good enough" to use, but they aren't today. Even the SOTA models barely code well, with Opus 4.5 being the first, good coding model.
That being said, I think it's absolutely imperative that we keep pushing local model performance. We need to continue to advance technology there and ensure that the model labs don't do regulatory capture in the name of "safety" (or anything else).
With the latest codex (weekly quota burn) fiasco I tried open weight alternatives for the first time. And tyeah... open weight models cant compete with likes of astra yet. But, my hope is that by the time I get my Mac studio at end of november an open weight models would have closed the gap (which i think is realistic at the speed of progress). Now its true a better gpt version will also be available then but it also seems the gap is shrinking with time so theres that.
Local models can be widely used as productive assets. Yes the infrastructure of SOTA API models is engineered specifically for you to be that utility, but the blanket statement that local isn't up to par is intensely short sighted. Billions of tokens per month on local pays for the hardware when compared to sota costs per month.
I believe they can currently be used productively for non-coding tasks (classification, light summary)... but they definitely are not even close to SOTA when it comes to software development.
Defining productivity is a use-case scenario, and a wildly generalized assumption for most people in this argument. Local infrastructure doesn't need to be sota for absolutely every single need for a dev lab, but it absolutely can be delivered with non-api frontier class models.
Just to be clear, I'm specifically talking about coding. I think local models can help with productivity today, just not coding.
I'm also a huge fan of local models and think it's absolutely imperative that they continue to advance so we can move off of the Anthropic/OpenAI hosted models. It's important to accurately asses where we are in that journey though.
Astra-Ultra? Even the largest open model to date (Kimi K3) is nowhere close to Astra level, and it will be quite slow even on the highest-spec M5 Ultra, with achievable speeds of about 0.5 tok/s at most due to having to stream weights from SSD (~13 GB/s on the highest storage capacity M5 Max machines so far). This is OK for doing simple Q&A in the background but it's far from a genuine coding experience. You'd have to test batching of multiple thinking streams in order to try and raise overall tok/s via layer-wise reuse of the streamed weights (and this is where the "Ultra" part sort of becomes relevant; Kimi series models have good support for agent swarms) but this would decrease single-session performance even further. It would only be usable for background jobs, though the hardware would then have a chance of paying for itself if it was fully used on a 24/7 basis.
Ehh, the actual elephant in the room is:
"why bother with local AI at all when you can lease a GPU for $5/hr?"
To which the answer is you shouldn't bother, unless you have a bunch of money to throw at hobby projects.
there are lots of people with very expensive hobbies, see sailboat racing for example.
"Lots" of people are billionaires too, just like "lots" of people like to watch TV in their free time.
While I know it's not apples to apples, the target comparison right now is 2x DGX Sparks. Similar price, 256gb. The conversation has focused on memory bandwidth vs. compute in agentic loops, so for most people the raw numbers will mean less than the "time per task" in coding benchmarks.
This is a great article and bodes well for the M5, but we should expect more like this comparing to other platforms before we truly understand where it fits.
Speed vs task-completion is a new conversation and a great point. Whereas the cost to compute doesn't exist in a vacuum, making mistakes costs less, is easier to maintain with granularity and a whole host of other factors when you own the lab.
Imagine spending a trillion dollars on data centers and then reading this article. Nightmare fuel for OpenAI
For 99.99% of people, spending 15 grand on a Mac Studio just to run Qwen 3.8 locally is a non starter.
It's not the M5 Ultra itself, but the M7s or M9s that will do the damage.
99% of people will use whatever AI is free. The sophisticated, heavy users that are willing and able to pay a lot of money the ones that will be interested in controlling their inference bills.
Today, the sweet spot where an M5 Ultra makes sense is tiny. But we might expect that to grow a lot.
Anthropic is reporting 100 Billion ARR.
Even if you could get a frontier model, you would not be able to run it on any Mac. So speculating on what M7 or M9 will achieve in 5 years (if we even still exist) seems pointless.
Do you need a frontier model to write emails, check your calendar, search the web?
I don’t think Apple is going to lie down and cede AI to the cloud.
at 15 tok/s
Yes, but in 7 years?
In 7 years we will probably all be living under ground fighting skynet with plasma rifles made out of old microwave parts
You will. Some of us are going to be already ground into dust that the microwave parts are made out of. Others will be locked into our communism cubes with our daily allotment of entertainment and sustinece. let out into the sunlight for only 30 minutes per day.
This dream machine costs over $15k (not including the Apple Studio Display)? Nah, that dream is SO OUT OF TOUCH!
If you are buying expensive hardware to run LLMs "on your own machine" you will soon find your ladder is on the wrong wall.
Ordered one for OpenFOAM. Excited for it. Will be nice to not have my laptop running CFD 24 hours a day, but my M1 Max is currently my fastest machine… I’m expecting about 3.5x from the M5 Ultra.
I think Apple is really sleeping on making this run a Linux server. These things are very capable and draw very little wattage when idle. It would make an excellent homelab device, but MacOS currently holds it back in this regard.
https://mac.getutm.app/
My mac is 5 years old. I don't think I can comfortably buy a new one right now. It has a 16GB unified RAM. Honestly that would be enough for so many local models that I want to use but can't use. Because RAM usage (even with literally every single user installed app quit/stopped) the RAM usage is very high that I can barely safely get 6-7 GB (I am supposed to get ~10 GB, but it goes up and down real fast!). That's a shame. If only I could install an alternative OS that uses very little amount of RAM :-)
When people benchmark MLX related quant models, they really need to publish numbers on benchmarks. You cannot take this as it is what you get of the original models. MLX uses pretty simple quantization methods so at lower bits without QAT, it is just not as good quality as llama.cpp ones.
At this point I think I will get the DGX gb300 workstation though I will wait a bit more for the cold season. It is double the price but at least is the real thing
The Year Of Local AI will be here no later than 2040, coinciding with the Year Of The Linux Desktop.
The year of linux on the Desktop was 26 years ago for me.
"a total cost of $0" Uhh ... how much is that hardware?
Another comment approximates at around USD$15k, so yeah, not zero.
Lol, try generation of images & videos on these, they ought to improve perf on Deep learning not just llms
What are good options to run local models nowadays? Something good for coding and personal assistant kind of things
I think we'll eventually get to the point where folks will have a local AI agent but I think people need to temper their expectations to a degree. You aren't going to have data center level tok/s from a box sitting under your desk and you don't need instantaneous responses for many workloads. Having a local agent that can execute tasks over a couple days with your supervision that might otherwise take you weeks is perfectly acceptable.
However I also think that Agentic AI is very much not an out-of-the-box solution, local or otherwise, and it takes a high level of technical knowledge to create an effective AI agent. And there's a problem now where most orchestration is fixed on what models are used for what tasks with no ability to weight constraints like cost, speed, and security.
Fantastic review, this is how it should be done with local AI.
i can buy a house with that amount of money. it used to be car money.