> peak power consumption is about 200W for the Mac Studio with M3 Ultra, and 300W for the Mac Studio with M5 Ultra — a fraction of the power used by a common toaster.
So, Meta will spend millions/billions on building open-weight models and then people will send Apple $20,000 to run them. Seems like a good deal for Apple.
Nvidia is also selling the shovle but they train their own LLM and by doing so they can fine tune their CUDA/Chip to have a better shovel than others.
Nvidia's own model does not have to beat OpenAI or any SOTA model. Nvidia just need to build up the in house expert and knowledge to make a better chip since hyperscalers and none hyperscaler like Meta is building their own AI chip now.
Apple can chose not to train their own model to learn how to sell a better shovel.
I've found some strange guardrails with the Foundation models. Like blocking describing of any photo with a single woman in it. It does this relatively consistently but with a different subset every time. Like out of 10 photos in the set, it's skips 6 then the next run, it skips 3. Have you seen guardrail issues like this?
I don't want any of these features and yet they've been forced upon me and are now getting in the way of me actually using my computer how I want. Games and heavy applications slow down and stutter because of "AI" background processes, even with App Tamer, my machine runs too hot to even touch when just idling. Compared to when I got my M2 in '23 and thought, "This is the fastest and snappiest machine I've ever used", my experience with an M5 and the latest macOS version is nothing short of an anti-consumer hellscape.
Interesting, this has not been my experience at all. I have an M5 machine and I'm not even aware of what AI features have been added, and it's streets ahead of my old M1 Pro in speed.
Well, the moat is that cpu/npu development is very hard, requires talented engineers, and requires access to cutting edge fab process, all of which in turn costs a lot of $$$. So, sure, “anyone” who has all that can definitely breach Apple’s moat.
Apples largest competitive advantage has always been and always will be - trust combined with epic marketing to make the vast majority of people aware of it.
Trust of quality, privacy w.r.t data etc.
I’m sorry but I’ve seen so many bad takes re. Apple over the years. Many of you think you are so smart on topics outside of your domain - but you’re not.
What exactly about the parent comment is wrong? Apple has always been good at making hardware that people are willing to pay a premium to use. I can easily imagine them playing a similar role in local LLMs to the role they played in smartphones in the 2010s.
apple is pretty well set up for this because they do all their hardware in house. if they wanted to, they could pivot in this direction. Ternus is also a hardware guy.
I think the bet is that it will get bigger, both at the consumer level, and with orgs eventually discovering ready-made solutions for having their own on-prem inference.
100,000 units a year would be $2 billion annually.
And usually, most of a company's revenue comes from the heaviest spenders.
So figure there are some out there who will by tens of these because they have some "idea to change the world". (Whether true or not.) Some startups, etc.
I'd wager, across planet earth right now, it should be trivial to push 100,000 units per year. Probably closer to 500,000 units per year. So that's an easy $10 billion a year.
$2-10 billion is meaningful for any company. Even one as large is Apple.
Especially if it's a first step towards a stronger foothold in the AI inference space.
They can then expand down market to grow the pie further.
For example, with truly dedicated inference boxes.
And given the demand for $7,000 Sparks and sold out $5,000 GPUs, the market for these $20,000 machines is likely much higher than my 100,000 unit lower bound. (Earth is a big place. It doesn't take much excitement to generate massive sales in a space with currently insatiable demand.)
They have the best strategy of bigtechco: let others figure out the squishier layers while they focus on selling the best hardware. Privacy has incredible value to a lot of individuals and industries. $20-80k in hardware that *appreciates* the moment you buy it is a reasonable investment for a lot of SMBs, and still less than a single Nvidia rig. Privacy pays.
If we're being frank here, Apple's AI strategy lags behind Nvidia, AMD and even Google. Nvidia has CUDA, AMD ships CDNA, Google has the TPU family, and Apple... well, Apple discontinued the rackmount Mac Pro, forfeit the HPC ARM datacenter to Nvidia's Grace CPUs, refuses to sign CUDA drivers, and spent half a decade designing raster GPUs instead of GPGPUs.
All of their competition has real inference and training stacks. Apple ships an iGPU and a smartphone NPU, they're reliant on Nvidia and Google hardware to deliver Apple Intelligence.
idk but naive me sees the article say the model needed 158GB of RAM so I guess 10x 16GB GPUs? Apparently retail price is $1k per 5080 16GB but they're actually selling at $1.5k so I guess $15k for just the GPUs.
158GB number is for Mistral Medium 3.5 128B model, Muse Spark is way larger.
Muse Spark Open would require more than 1.6TB of RAM in its native precision (BF16) for the model parameters alone. In a 4-bit quantized (NVFP4) format, the model memory footprint is lowered to about 493GB.
In any case, I'd much rather run GLM 5.3 Flash on that sort of hardware. Better fit for the performance profile, and will probably beat Muse Spark Open in most uses.
I think they're sandbagging Qwen pretty heavily by benchmarking Meta's model at 4-bit, against Qwen only at native BF16 (and nothing else shown). That's not a quant most users would choose.
I get that this is cross-brand marketing for Apple's partner, Meta; but really Muse Spark isn't even in the top *ten* models I'd consider if I had a 512 GB M5 in my paws. Or a cluster of them.
FWIW, single M5 Ultra with 256GiB runs DeepSeek 4.1 Flash 0910 just fine (~3200 tok/s prefill, ~45 tok/s). M5 Max w/ 128GiB runs GLM 5.3 Flash also fine.
I also benchmarked both on quality benchmarks (MathArena Apex / GPQA Diamond etc) to make sure the quality is what you get from the official numbers.
In this post, we show how to run the latest frontier-class model — Muse Spark Open, with nearly a trillion parameters — right from your desk, using the latest version of MLX, released today. On a cluster of four Mac Studios, Muse Spark Open generates over 36 tokens per second (see Figure 1), well above average human reading speed. Beginning on a single Mac Studio (up to 512GB of unified memory), then scaling out across multiple Macs with the RDMA over Thunderbolt driver, below we share an overview of the range of models you can now run on Mac Studio.
Yes, I'm very glad they're doing this, but as of 2026, it's a bit academic. Pricing for the 512GB Mac Studio hasn't even been announced yet, but the 256GB version starts at $9499. So when you start talking about four of those puppies, it's substantial.
Someone else will do the math on how many years it would take to spend $40k on Muse Spark Open via OpenRouter.
But there are definitely people out there that will need to run these models locally (contractual or legal requirements). By validating a Mac Studio as an option for running local models, it might be the more appealing option.
For most of us though, you're right -- the OpenRouter option is far more cost effective.
It seems Apple figured they could sell more Mac Studios on the basis of frontier local AI models, profit to be made while users haven't figured out the math yet.
Don't get me wrong; in the long term I see the price of local inference falling as their chips continue to get 15-20% faster every year, and meanwhile the "intelligence per billion parameter" chart is rising. There will be a magic point where they cross, and "good enough" reaches your phone. That's when Apple starts to drink Nvidia's milkshake.
It’s not about the price of Apple products falling—I was commenting that the price of inference will continue to fall (chips getting faster). If “good enough” inference is in your pocket, this benefits Apple and harms Nvidia.
You can't really compare API prices though. The cards that they are running frontier models on in the datacenter can service hundreds of simultaneous requests and are at nearly 100% utilization. This thing would sit on your desk and be used to service a single user for a couple hours per day.
It was only like $14k new. Why are you people acting like nobody has this stuff? People were already building clusters when that thing first came out. The research isn't bunk just because the refresh hasn't landed yet, my god.
Maybe you (or i?) missed the point - it wasn't that nobody has those machines, it was that at $40k for 4x the 256GB versions is a substantial price, and so the parent parent was suggesting that it would be compared against OpenRouter. Ie how long would it take someone to spend $40k worth of credits on OpenRouter?
I've had the same thought on self hosting for the past 2 years. Eg i could invest $10k into hardware or spend those on credits. I chose credits, because by the time i spend 10k i suspect the market will have dramatically shifted in requirements, compute, etc.
I find it bizarre of Apple to do anything publicly in partnership with Meta. I always thought the reasons Apple snubbed Facebook were out of values and principles not business. That could be wrong. Now that Meta has a line of business that Apple can finally charge a platform tax on there is new reason to start sleeping with them.
> I always thought the reasons Apple snubbed Facebook were out of values and principles not business.
Apple is showing a masterclass of how to tow the line perfectly. Or more candily "have the cake and eat it too".
If you ask anyone, what's one company that'll stand up to government requests, most reasonable people would say Apple. Yet, Apple is extremely strategic when it comes to China. It thrives in that country, without being associated with CCP.
Same here - they are happy to partner with Meta, without getting any of the bad "aura".
To be clear, all companies do this just that Apple alone has a successful strategy here. Tim Cook personally gifted the current adminstration with a gold plaque, something none of the other "dirty" CEOs did. Such a public gesture, yet nobody would think that Apple is in bed with the current administration. Its marvelous to see this happen in real time.
I worked at a company where one of the pieces of hardware we were building at the time used a PowerPC CPU (this is back in ~2007). We had a rack of Apple xServe for build and test. The xServe was nice pieces of hardware compared to the x86 kit we had, more like Sun gear in their attention to detail and ease of use.
At my current company we compile for x86 and devs can do a full build and test environment on their laptops (which are 95% MacBook Pros) but it is unclear what will happen when Rosetta support is pulled. Now the interesting thing is we plan to move our stuff over to target the AWS Graviton (and other cloud ARM cores) as those instances are cheaper for generally the same performance (for what was are doing anyway) so the point will be moot. There will be some crossover point where having build and test servers in a rack will be cheaper than doing it in the cloud. At that point having Apple ARM based servers would be a win assuming we can build a binary that works for Graviton, et.al.
It's a good article which showcases direction of travel, i am sure mac studio will get more and more powerful in coming generations but for now it's way to expensive for local setup.
yeah especially because it's not very easy to configure everything sensibly for a multi-use case setup, they really need better mac containers on the mac itself
While right now Mac Studio is not a budget device and it certainly costs more than the subsidised subscription, the promise that one day we can run these models on commodity hardware is exciting.
> In this post, we will show how to run the latest frontier-class model - Muse Spark Open, with nearly a trillion parameters - right from your desk, using the latest version of MLX, released today.
You will get
> In this post, we will show how to run the latest frontier-class model – Muse Spark Open, with nearly a trillion parameters – right from your desk, using the latest version of MLX, released today.
Not everything with em dashes is the product of generative AI... some of it is the word processor correctly following the rules of grammar and punctuation.
If used correctly, I find the em dash one of the most elegant and reader-friendly ways to structure certain parts of a text — that it got such a bad reputation lately really is a shame!
> peak power consumption is about 200W for the Mac Studio with M3 Ultra, and 300W for the Mac Studio with M5 Ultra — a fraction of the power used by a common toaster.
(Emphasis added.)
I run my toaster for hours on end too. Love my sandwiches.
So, Meta will spend millions/billions on building open-weight models and then people will send Apple $20,000 to run them. Seems like a good deal for Apple.
Exactly why does Apple need to do any model training/development? They are out here selling the shovels.
Nvidia is also selling the shovle but they train their own LLM and by doing so they can fine tune their CUDA/Chip to have a better shovel than others.
Nvidia's own model does not have to beat OpenAI or any SOTA model. Nvidia just need to build up the in house expert and knowledge to make a better chip since hyperscalers and none hyperscaler like Meta is building their own AI chip now.
Apple can chose not to train their own model to learn how to sell a better shovel.
Apple’s Photos ecosystem is privacy oriented and as a result some (now) baseline features require the use of local models.
The Foundation models keep a surprisingly decent set of data available for fast filtering by a local LLM.
I’ve found a slower, second pass, is needed by Qwen to clean it up. (I am running on the new Mac Mini m5 pro 64gb.)
I’ve run my own tests using a date range, a short description of a red cedar planter box build I did last year, and a set of known correct photos.
There is some cool stuff you can do without sending your family’s pics up to any of the frontier model providers.
I've found some strange guardrails with the Foundation models. Like blocking describing of any photo with a single woman in it. It does this relatively consistently but with a different subset every time. Like out of 10 photos in the set, it's skips 6 then the next run, it skips 3. Have you seen guardrail issues like this?
I switched to Gemma 4 instead.
same i am using new embeddinggemma. seems far more capable than foundation models.
I don't want any of these features and yet they've been forced upon me and are now getting in the way of me actually using my computer how I want. Games and heavy applications slow down and stutter because of "AI" background processes, even with App Tamer, my machine runs too hot to even touch when just idling. Compared to when I got my M2 in '23 and thought, "This is the fastest and snappiest machine I've ever used", my experience with an M5 and the latest macOS version is nothing short of an anti-consumer hellscape.
Interesting, this has not been my experience at all. I have an M5 machine and I'm not even aware of what AI features have been added, and it's streets ahead of my old M1 Pro in speed.
Exactly my thought as I read it.
They make and sell the hardware that will actually make AI something everyone can used without needing to rely on OpenAI/Anthropic et al.
Apple has never had a moat, anyone else could have built this sort of hardware but they haven't.
So they don't care which model wins, because it will always move out to the edge of the network, and they own a large chunk of that edge.
Well, the moat is that cpu/npu development is very hard, requires talented engineers, and requires access to cutting edge fab process, all of which in turn costs a lot of $$$. So, sure, “anyone” who has all that can definitely breach Apple’s moat.
Ok.. and why specifically do people work at Apple?
Hint: money is not the only reason.
Jeez.
Ermmm what?
Apples largest competitive advantage has always been and always will be - trust combined with epic marketing to make the vast majority of people aware of it.
Trust of quality, privacy w.r.t data etc.
I’m sorry but I’ve seen so many bad takes re. Apple over the years. Many of you think you are so smart on topics outside of your domain - but you’re not.
What exactly about the parent comment is wrong? Apple has always been good at making hardware that people are willing to pay a premium to use. I can easily imagine them playing a similar role in local LLMs to the role they played in smartphones in the 2010s.
> Apples largest competitive advantage has always been and always will be - trust.
Lmao
> Many of you think you are so smart on topics outside of your domain
Turn the mirror around
apple is pretty well set up for this because they do all their hardware in house. if they wanted to, they could pivot in this direction. Ternus is also a hardware guy.
How big is the local LLM market? I have trouble imagining Apple will move enough units to make $20k/ea worth it for them as a business segment.
They have been sold out of a lot of units, due to people developing AI, so apparently, it is significant.
I think the bet is that it will get bigger, both at the consumer level, and with orgs eventually discovering ready-made solutions for having their own on-prem inference.
100,000 units a year would be $2 billion annually.
And usually, most of a company's revenue comes from the heaviest spenders.
So figure there are some out there who will by tens of these because they have some "idea to change the world". (Whether true or not.) Some startups, etc.
I'd wager, across planet earth right now, it should be trivial to push 100,000 units per year. Probably closer to 500,000 units per year. So that's an easy $10 billion a year.
$2-10 billion is meaningful for any company. Even one as large is Apple.
Especially if it's a first step towards a stronger foothold in the AI inference space.
They can then expand down market to grow the pie further.
For example, with truly dedicated inference boxes.
And given the demand for $7,000 Sparks and sold out $5,000 GPUs, the market for these $20,000 machines is likely much higher than my 100,000 unit lower bound. (Earth is a big place. It doesn't take much excitement to generate massive sales in a space with currently insatiable demand.)
They have the best strategy of bigtechco: let others figure out the squishier layers while they focus on selling the best hardware. Privacy has incredible value to a lot of individuals and industries. $20-80k in hardware that *appreciates* the moment you buy it is a reasonable investment for a lot of SMBs, and still less than a single Nvidia rig. Privacy pays.
If we're being frank here, Apple's AI strategy lags behind Nvidia, AMD and even Google. Nvidia has CUDA, AMD ships CDNA, Google has the TPU family, and Apple... well, Apple discontinued the rackmount Mac Pro, forfeit the HPC ARM datacenter to Nvidia's Grace CPUs, refuses to sign CUDA drivers, and spent half a decade designing raster GPUs instead of GPGPUs.
All of their competition has real inference and training stacks. Apple ships an iGPU and a smartphone NPU, they're reliant on Nvidia and Google hardware to deliver Apple Intelligence.
How many 5080 16 Gig cards would you have to buy to run the same model, and what would that cost?
idk but naive me sees the article say the model needed 158GB of RAM so I guess 10x 16GB GPUs? Apparently retail price is $1k per 5080 16GB but they're actually selling at $1.5k so I guess $15k for just the GPUs.
158GB number is for Mistral Medium 3.5 128B model, Muse Spark is way larger.
So about 24gb/expert at NVFP4. If you compressed the KV cache, you might be able to get away with 2x 24GB cards and stream the experts.
If you're happy with generation speed in 5 t/s range, sure -- PCIe 5.0 x16 is only 64GB/s.
I never said it would be fast!
In any case, I'd much rather run GLM 5.3 Flash on that sort of hardware. Better fit for the performance profile, and will probably beat Muse Spark Open in most uses.
It won't be free for Apple to use Meta's model in the future.
Both companies are profitable and AI is not their main line of business. This affects pure AI companies.
As open weight models continue to match frontier models, the race to $0 will continue to benefit supporters of running open weight models locally.
I think they're sandbagging Qwen pretty heavily by benchmarking Meta's model at 4-bit, against Qwen only at native BF16 (and nothing else shown). That's not a quant most users would choose.
I get that this is cross-brand marketing for Apple's partner, Meta; but really Muse Spark isn't even in the top *ten* models I'd consider if I had a 512 GB M5 in my paws. Or a cluster of them.
FWIW, single M5 Ultra with 256GiB runs DeepSeek 4.1 Flash 0910 just fine (~3200 tok/s prefill, ~45 tok/s). M5 Max w/ 128GiB runs GLM 5.3 Flash also fine.
I also benchmarked both on quality benchmarks (MathArena Apex / GPQA Diamond etc) to make sure the quality is what you get from the official numbers.
Not exactly your average "desk", but still neat.
Yes, I'm very glad they're doing this, but as of 2026, it's a bit academic. Pricing for the 512GB Mac Studio hasn't even been announced yet, but the 256GB version starts at $9499. So when you start talking about four of those puppies, it's substantial.
Someone else will do the math on how many years it would take to spend $40k on Muse Spark Open via OpenRouter.
But there are definitely people out there that will need to run these models locally (contractual or legal requirements). By validating a Mac Studio as an option for running local models, it might be the more appealing option.
For most of us though, you're right -- the OpenRouter option is far more cost effective.
It seems Apple figured they could sell more Mac Studios on the basis of frontier local AI models, profit to be made while users haven't figured out the math yet.
Don't get me wrong; in the long term I see the price of local inference falling as their chips continue to get 15-20% faster every year, and meanwhile the "intelligence per billion parameter" chart is rising. There will be a magic point where they cross, and "good enough" reaches your phone. That's when Apple starts to drink Nvidia's milkshake.
Apple and NVIDIA prices are not going to fall.
Their competitors will. Anyone locked into these two ecosystems are willingly going along with it.
It’s not about the price of Apple products falling—I was commenting that the price of inference will continue to fall (chips getting faster). If “good enough” inference is in your pocket, this benefits Apple and harms Nvidia.
Puts into perspective how cheap the frontier models are via subs or even API.
You can't really compare API prices though. The cards that they are running frontier models on in the datacenter can service hundreds of simultaneous requests and are at nearly 100% utilization. This thing would sit on your desk and be used to service a single user for a couple hours per day.
There was the M3 Ultra 512GB Mac Studio, though.
…which is in the region of €20k on eBay, so the calculus doesn’t really shift…
It was only like $14k new. Why are you people acting like nobody has this stuff? People were already building clusters when that thing first came out. The research isn't bunk just because the refresh hasn't landed yet, my god.
Maybe you (or i?) missed the point - it wasn't that nobody has those machines, it was that at $40k for 4x the 256GB versions is a substantial price, and so the parent parent was suggesting that it would be compared against OpenRouter. Ie how long would it take someone to spend $40k worth of credits on OpenRouter?
I've had the same thought on self hosting for the past 2 years. Eg i could invest $10k into hardware or spend those on credits. I chose credits, because by the time i spend 10k i suspect the market will have dramatically shifted in requirements, compute, etc.
Hilarious how they managed to put out an entire press release about Meta’s technology without mention Meta other than what is in the code references.
For the press, you had the release from meta here: https://research.meta.ai/blog/muse-spark-open
...and Zuck and Ternus high-fiving each other here: https://x.com/finkd/status/2110778025629347999
"We're excited for mommy and daddy to be playing nice again," Meta AI chief Alexandr Wang told Axios here: https://www.axios.com/2026/10/15/apple-and-meta-bury-the-hat...
It’s not a press release and not about Meta’s technology.
The title (“Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio”) doesn’t even mention Muse.
It’s more like an advertorial for their tech.
Still weird, and against brand, for them to lead with this particular model from that particular company. The groans are still earned.
I find it bizarre of Apple to do anything publicly in partnership with Meta. I always thought the reasons Apple snubbed Facebook were out of values and principles not business. That could be wrong. Now that Meta has a line of business that Apple can finally charge a platform tax on there is new reason to start sleeping with them.
> I always thought the reasons Apple snubbed Facebook were out of values and principles not business.
Apple is showing a masterclass of how to tow the line perfectly. Or more candily "have the cake and eat it too".
If you ask anyone, what's one company that'll stand up to government requests, most reasonable people would say Apple. Yet, Apple is extremely strategic when it comes to China. It thrives in that country, without being associated with CCP.
Same here - they are happy to partner with Meta, without getting any of the bad "aura".
To be clear, all companies do this just that Apple alone has a successful strategy here. Tim Cook personally gifted the current adminstration with a gold plaque, something none of the other "dirty" CEOs did. Such a public gesture, yet nobody would think that Apple is in bed with the current administration. Its marvelous to see this happen in real time.
Apple also was one of the largest contributors to the ballroom. They're not maintaining much of a "moat" between the company and the administration.
Its still weird though.
The datacenters are not full of Mac studios but Nvidia hardware.
They are known for macbooks.
And they are not willing to build proper apple servers for all the build systems people have.
They did, and few people bought them, so they stopped.
I worked at a company where one of the pieces of hardware we were building at the time used a PowerPC CPU (this is back in ~2007). We had a rack of Apple xServe for build and test. The xServe was nice pieces of hardware compared to the x86 kit we had, more like Sun gear in their attention to detail and ease of use.
At my current company we compile for x86 and devs can do a full build and test environment on their laptops (which are 95% MacBook Pros) but it is unclear what will happen when Rosetta support is pulled. Now the interesting thing is we plan to move our stuff over to target the AWS Graviton (and other cloud ARM cores) as those instances are cheaper for generally the same performance (for what was are doing anyway) so the point will be moot. There will be some crossover point where having build and test servers in a rack will be cheaper than doing it in the cloud. At that point having Apple ARM based servers would be a win assuming we can build a binary that works for Graviton, et.al.
They don’t lead with it. The title of the actual article is “Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio”.
The submitter changed it to “Apple ML Research: Muse Spark Open with MLX on Mac Studio”
Second paragraph: "In this post, we show how to run the latest frontier-class model — Muse Spark Open, with nearly a trillion parameters..."
Anyone know why they went with Meta models?
It's a good article which showcases direction of travel, i am sure mac studio will get more and more powerful in coming generations but for now it's way to expensive for local setup.
yeah especially because it's not very easy to configure everything sensibly for a multi-use case setup, they really need better mac containers on the mac itself
A 800B model that performs way worse than other 300B ones. Well...
Why buy one 256 GB Studio, when you could buy two 512 GB's for lower-quality output?
Local LLM's really sharpen your senses to model size. You could fit a car in the price difference!
What are you using to determine performance?
Most people are probably looking at: https://artificialanalysis.ai/models/comparisons?compare=mus...
This is still a welcome model, even if I won't be running it at home
I am super excited about this.
While right now Mac Studio is not a budget device and it certainly costs more than the subsidised subscription, the promise that one day we can run these models on commodity hardware is exciting.
Preparing to upsell those very expensive Macs coming later this month, are we.
I -- love -- an article -- full of -- em -- dashes.
If you type the following into Microsoft Word:
> In this post, we will show how to run the latest frontier-class model - Muse Spark Open, with nearly a trillion parameters - right from your desk, using the latest version of MLX, released today.
You will get
> In this post, we will show how to run the latest frontier-class model – Muse Spark Open, with nearly a trillion parameters – right from your desk, using the latest version of MLX, released today.
Not everything with em dashes is the product of generative AI... some of it is the word processor correctly following the rules of grammar and punctuation.
If used correctly, I find the em dash one of the most elegant and reader-friendly ways to structure certain parts of a text — that it got such a bad reputation lately really is a shame!
Em dashes—love ’em!—but they do. not. need. personal. space.