React on Rails HN RSC Demo

ks2048 5d ago

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.

Juvination 5d ago

Exactly why does Apple need to do any model training/development? They are out here selling the shovels.

bredren 5d ago

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.

cj00 5d ago

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.

dominotw 5d ago

same i am using new embeddinggemma. seems far more capable than foundation models.

blacksmith_tb 4d ago

I am going to assume that means "a women by herself" and not that it can magically detect her marital status before refusing to continue (I suppose it could be laser-focused on wedding rings...)

sneak 4d ago

No. Apple’s Photos ecosystem uploads all Photos effectively unencrypted to Apple in a way that is readable by Apple, allowing Apple to do server-side analysis of all content. This content is also provided to US or Chinese federal authorities without due process or warrants.

Nothing about iCloud Photos is private.

isomorphic 4d ago

With Advanced Data Protection enabled (it's not available globally but it is in the US), Photos is E2EE.

sneak 4d ago

Even with ADP on, it uploads the unique hashes of the PLAINTEXT of every photo to Apple, so they know who had what unique files, when, and who had them next, and when, etc. You can’t receive or transmit an original Winnie The Pooh meme between two phones with ADP on without Apple and the FBI knowing who had it first and when the second person received it.

Also insofar as ADP is off by default the whole system can NOT be said to be “privacy oriented”, which was the claim I was refuting.

Apple, for the most part, does not care about the privacy of their users. ADP would not be built today, and it will probably eventually be quietly removed (like it was in the UK), because ~nobody uses it.

soulofmischief 5d ago

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.

papa_bear 5d ago

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.

soulofmischief 4d ago

peep the comment I just made to the adjacent commenter, that has been my experience. I basically have to run my fans at full speed all day while working now in order to type on my computer (when I'm not using my magic keyboard). It is like a huge problem, not even just a small one but like something that is affecting me negatively every single day, increasingly making me want to break this machine in half and disappear into the woods.

sneak 4d ago

The M1 used to be adequate in speed. Then Apple kept adding always-on cloud daemons that run even if you don’t use iCloud.

doug_durham 5d ago

My Macs run cool and silently. I've never observed a background "AI" process consuming resources.

soulofmischief 4d ago

It's not just AI processes. But things like mediaanalysisd, duetexpertd, etc. even the freaking menu bar agent, all regularly pound my computer both when idling and when in use.

I recently installed App Tamer which is helping a bit, but it doesn't allow me to tame everything. I just checked now and duetexpertd was trying to use nearly two full cores. I have all iCloud integrations turned off, I do not use iCloud at all, and I do not have an iPhone, so I am flabbergasted that this iCloud desktop-phone syncing process runs at such high core usage in the background with extreme regularity. Proof: https://imgur.com/VkBh9bi

Unfortunately, this does nothing for services such as WindowServer which now often takes up 50-100% of a core just on its own.

I have reinstalled and upgraded to the latest version. This MacBook is only a few months old.

While typing this message and literally just having open Codex with one running task that is not doing anything substantial, and Firefox, with no high CPU usage from any tab, my laptop got so hot that I had to use another installed program, Stats, to manually bump up the fan since macOS also has absolutely garbage native system fan controls.

I have fibromyalgia and seem to be sensitive to hot keys; my Thinkpad regularly getting to hot to touch with even light usage is EXACTLY why I went and bought a MacBook one morning in '23. One morning I touched my keyboard and realized I was tired of singeing the nerves on my fingertips and having them hurt for hours. So now in '26, finding myself in the same exact position but with a much more expensive computer, I feel like I am living in a corporate hellscape nightmare.

One thing I have noticed is all of my cores can sit idle except a single one and that will still heat up the machine a lot. If the hardware is designed as such, the software MUST be designed to work around this and not blast a core for 10 minutes straight every time one of a dozen or two misbehaving system processes wants to complete a background task. If the software is not designed for this, the hardware is incomplete and faulty. This would be true if it was shipped as firmware; software doesn't get a free pass here.

But now Pages and Numbers recommend document titles for me, so.... worth it?

roughly 3d ago

Huh, I hadn't seen App Tamer before, that's quite a neat bit of kit. St. Clair does serious work...

soulofmischief 3d ago

I'm glad I found it. Jettison looks like a nice tool as well. It's working great so far, just does what I need it to do. Fits in nicely with Stats, which I use for fan control and monitoring power consumption. I also found https://github.com/Digital-Shane/OpenTamer but figured App Tamer was still the more mature option.

Angostura 4d ago

I’m on an M1 workout these issues. How are you isolating them to be AI related?

soulofmischief 4d ago

medianalysisd for example, which frequently pummels 1-2 cores of my machine for long stretches of time, runs local on-device models to analyze the media on your machine.

rswail 5d ago

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.

intrasight 4d ago

> Apple has never had a moat, anyone else could have built this sort of hardware but they haven't.

I think both points are very incorrect. Apple has a massive moat. And Apple is making hardware (specifically, CPUs) that are two years ahead.

rswail 2d ago

That "moat" is being reduced (again, compared to previous Apple history) as other companies catch up. Apple has continually cannibalized their own hardware to keep ahead of others.

Unlike the "pure" AI players, who rely on NVIDIA and others, Apple is relying on its own capabilities, vertically organized, to maintain its lead in consumer space.

They are using the "cloud" to again, vertically integrated, to preserve their "privacy first" consumer marketing, their hardware to maintain their "premium product" lead, and they are usually second to a market after a first-mover advantage has been reduced by ongoing development.

drob518 5d ago

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.

rswail 2d ago

Do Microsoft, Google and others not have access to all of those items?

uurue1 5d ago

Ok.. and why specifically do people work at Apple?

Hint: money is not the only reason.

Jeez.

drob518 3d ago

I’m at a loss as to why you think I said that people would only work at Apple for money.

sneak 4d ago

Most of the smart people I know have left Apple because over the last ten or fifteen years, all of the reasons other than money have gone away, and they don’t pay well.

uurue1 5d ago

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.

Nevermark 4d ago

Don’t underestimate how much of their success is due to unusual internal advantages over competitors.

They spent decades learning to eliminate a 1000 headaches for themselves, and many for their customers, by accumulating vertical integration:

The co-design of their own comms, processors, devices and OS variants. Their own integrated supply chain and sales channels.

The better the results, the more invisible or clock-like their delivery.

Very easy to overlook. But nobody can match them on their terms.

(Nintendo also applied a unique set of internal advantages to generations of successful devices: vertically integrating humble components, highly scoped software, a strong theory of fun, and deep branding with novel polished interfaces.)

digitaltrees 4d ago

This post could be taught in business school. Kudos.

Apple is the only company that actually listened to Alan Kay when he said if you care about software, build hardware.

Danox 4d ago

??? Apple’s advantage over their competition is the fact that they do hardware, Apple Silicon and they do operating systems Apple OS very very well, in comparison to the present competition in the marketplace.

They are the last company that’s vertical from the 1980s. Unfortunately, all the other companies (SGI, Sun, Digital etc..) died in the face of the Wintel colossus.

Marketing is irrelevant. If you don’t have something to sell that is actually any good see the Microsoft Surface fiasco and its lack of an impact in the marketplace.

fasterik 5d ago

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 personal computing in the 1980s or smartphones in the 2010s.

oakejp12 5d ago

> 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

MangoCoffee 5d ago

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.

jtbaker 4d ago

> They are out here selling the shovels

just wish their containerization / linux compat wasn't so nerfed. Just moved my homelab's inference engine over to a GB10 from a Mac MTP setup to have first class k8s support, as well as open up access to the broader CUDA ecosystem (cuOpt etc.).

tyre 4d ago

They’re not really. They make the things useful for two classes of users: Mac Studio for people running models locally and Siri backed by Gemini for voice interaction.

They’ll likely embed text models more deeply in the future, but the history of “Apple Intelligence” doesn’t give a lot of hope that they’ll nail it.

So they’re not selling shovels (a precursor to making the thing) but products and services to use it.

jorblumesea 5d ago

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.

neya 4d ago

That's just BS cope. They tried and miserably failed - and still are failing badly. Remember, they were the ones to push for AI with "Siri". Many firings later and 2 CEOs later we are still exactly in square one. The fucking thing can't even get who I want to call, let alone set an alarm without touching my phone FFS.

Danox 4d ago

What is there to cope about? Apple is the last vertical computer company left from the 1980s. Apple Silicon just gets better every year. They have also managed to replace several companies that well gave them problems over the years or said no to doing anything that involved moving into the future. i.e, Motorola of Schaumburg, Illinois, IBM and Intel one company’s out of business. and the other two are trying to hang on to past glories.

By the way, the new Apple folding phone appears to be a hit. a real hit the last test left is whether or not it is as good as it appeared to have been in the demonstrations. only time will tell when it gets into the hands of the general public.

pjmlp 4d ago

Lets not overlook the fact that Apple is only around due to sheer luck with NeXT reverse acquisition, and having Steve Jobs back on the steering wheel.

Nowadays they are basically the iDevices company, if macOS wasn't required for XCode, the worldwide sales would be much less.

neya 4d ago

None of what you said is relevant to my comment about Apple's failed AI strategy nor Siri.

Lio 4d ago

I have no interest in Siri or always-on AI assistants. I turn as much of that off as I can.

I do have an interest in local, private inference though and I am willing to pay for that.

Apple segmenting the market is actually a good strategy. Horses for courses, as they say.

neya 4d ago

> local, private inference

Yeah, sure, inside a closed-source OS communicating constantly to private servers on the internet. Sounds very private to me.

Lio 2d ago

As long as you can disconnect it from the network then it's more private than something running on a cloud server.

oulu2006 4d ago

Apple playing 3d chess here -- they can just absolutely kill it in making the HW available for everyone to run models at home, and will do a better job than Nvidia especially when M7 comes out which is specifically designed for AI workloads.

rvz 5d ago

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.

loeg 5d ago

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.

ijidak 5d ago

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.)

____tom____ 5d ago

They have been sold out of a lot of units, due to people developing AI, so apparently, it is significant.

Danox 4d ago

The only question you need to ask is whether or not Apple will do anything to design and engineer memory in-house, the money, and the capability is there that is the challenge. because you don’t need to select $20,000 computer if you get that under control by bringing it in-house.

In short, Apple needs to roll up there sleeves and get busy the three hyenas of memory and that nameless memory company in China will not sell you memory when you need it at a reasonable price. In the past Motorola said no, IBM said no, and Intel said no. This is another one of those times for Apple to adapt by designing and engineering around the competition.

y1n0 4d ago

The memory constraint isn’t design. It’s fab capacity. Apple has exactly 0 fabs. By the time they hired domain experts and built fabs and tuned to the point of reasonable yields, the crunch would have been over for 5 years, and apple would have wasted 20 billion with no return.

Lalabadie 5d ago

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.

throwaway27448 4d ago

Probably worth it just as an investment into their developer ecosystem. They could probably sell at a deep loss and still justify it.

reilly3000 5d ago

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.

aurareturn 4d ago

I disagree.

I think the iOS App Store is going to be in steep decline in a few years. People will generally have a few apps: ChatGPT, Instagram, messaging, and something like a Photos/Camera app.

Everything else will be done in your AI app.

I've been hooking up more and more data and plugins into ChatGPT/Claude. I only want to use it. I don't want to use other apps if I don't have to. For example, research, buy, plan. I prefer not to use other apps.

I think all it takes is for OpenAI to make a phone from the ground up for this and Apple is going to have serious competition.

bigyabai 5d ago

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.

AndrewKemendo 4d ago

Saying they are lagging assumes they are in the same race

They aren’t

Apple is a high net worth consumer technology integrator that focuses on highest coherence user experience. They pick the most efficient stable technology, integrate it and sell the package

Apple only needs to lead in precisely that they don’t need to lead in literally anything else

in fact if they tried to lead in those other things they would give up their integration capacity

bigyabai 4d ago

But they are. Nvidia ships GPGPU ARM SOCs to consumers; why can't Apple? They're both trillion-dollar companies, surely Apple isn't lacking in engineering talent.

The vertical integration argument relies on Apple having a good architecture for AI. First it was the ANE, which everyone (mistakenly) thought would differentiate Apple Silicon from other NPU platforms. Then everyone parroted the "unified memory" talking point without checking to see if professionals actually used it. Now people are pivoting to say the next-gen Apple GPUs will finally blow the last-gen GPGPUs out of the water. Which is good, because there are 4-year-old GPUs that beat M5 Ultra in compute efficiency.

I know that "Apple is behind" is a tough pill to swallow, but they literally have no analog to CUDA, CDNA or even TPU technology. Nvidia SOCs like Tegra prove that Apple could absolutely bring GPGPU tech to low-power hardware like the iPad and iOS without giving up any real features. Apple just refuses to invest in it.

AndrewKemendo 4d ago

I feel like you didn’t read my original post cause you’re just repeating the original claim that they need to compete in components market instead of consumer product market

GeekyBear 5d ago

How many 5080 16 Gig cards would you have to buy to run the same model, and what would that cost?

fluidcruft 5d ago

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.

petu 5d ago

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.

bigyabai 5d ago

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.

petu 5d ago

If you're happy with generation speed in 5 t/s range, sure -- PCIe 5.0 x16 is only 64GB/s.

binlog 4d ago

Now scale that up to Nvidia and you've understood the entire AI market

giancarlostoro 4d ago

Still cheaper than the equivalent Nvidia GPU setup btw.

pjmlp 4d ago

The NVIDIA GPU can be changed as new models become available, without throwing away the whole computer.

JKCalhoun 4d ago

And the Mac Studio can continue to be used as a regular computer. (Hilariously over-qualified, I suppose. But, no, I got a maxed out, 128 GB Mac Mini earlier in the year—when you still could—and I find it a snappy daily driver in my 'man cave' downstairs.)

pjmlp 4d ago

While failing to satisfy existing Mac Pro demographics, now slowly migrating to Windows and Linux worstations, for all kinds of professional cards they require.

Naturally there are those that will put up with a Studio and thunderbolt cable spaghetti.

giancarlostoro 4d ago

Thunderbolt still a thing? I coulda sworn it was all USB C now.

giancarlostoro 4d ago

The lifespan of those AI GPUs is like 5 years, I have macs from 2008 that still turn on and function.

pjmlp 4d ago

So what, my Timex 2068 from 1986 also turns on and functions, does not mean it is useful in 2026.

throwaway27448 4d ago

That is generally how hardware gets sold yes

newsclues 4d ago

I wonder how many units Meta bought to convince Apple to do a collaborative advertisement.

informal007 5d ago

It won't be free for Apple to use Meta's model in the future.

Danox 4d ago

Google, who is far more qualified, and is getting a measly 1 billion dollar per year refund on that $20 billion dollars they currently paying Apple for a default search position on Mac OS. Meta, whose software prowess is questionable, when compared to Google or Microsoft isn’t even in the game they are getting how should I put this Apples charity/tolerance which is all OpenAI and Google are currently getting.

One is getting a $1 billion refund from what they already pay Apple, and the other two are getting nothing from Apple.

Meta won’t be getting anything from Apple in the future, other than tolerance/charity.

chaostheory 4d ago

Even in the unlikely case that does happen, there are plenty of other open weight models like Qwen, GLM, and Mistral to name a few

peri-cl 5d ago

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.

dannyw 4d ago

They clearly list out the quants in Table 1.

Most of the following tables show relative speedups; not tok/s numbers.

I agree no home AI users should really use bf16^; but I see two defensible justifications: (1) showing that their new hardware's speedups scale to full precision too [crucially; NVIDIA is focusing more and more on lower quants], and (2) trying to sell to the crowd running IQ3XXS or something on 16GB.

^q8 ~= bf16 is a good general rule of thumb. And friends don't let friends quantize KV cache.

AceJohnny2 4d ago

> cross-brand marketing for Apple's partner, Meta

Partner!? I don't see where you get that from the article, and Apple & Facebook have traditionally not been best of friends (see everything iOS implemented to limit Facebook's overreach).

I'm guessing they chose Muse because 1. It's not Chinese so it'll look good for Trump's AI cronies (Apple is in the administration's spotlight) and 2. It's fairly recent and thus on people's mind.

peri-cl 4d ago

I mean that they're doing joint marketing on this particular thing. Muse Spark marketing features Mac Studios quite prominently, and Apple's material spotlights Muse Spark.

mattmaroon 4d ago

Eh, Apple spent years suing Samsung while also buying parts from them for their products. You can be a partner and an enemy at the same time at the high levels of the game.

solarkraft 3d ago

For what it’s worth, Muse Glimmer is by far the most usable model on my 32GB M1 Max. It gets stuff done very competently while Qwen never gets done reasoning.

liuliu 5d ago

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.

TomGarden 4d ago

GLM 5.3 Flash on 128GB Ram would have to be 2-bit right?

I'm no expert but it seems the consensus is you get a suboptimal experience under 4-bit. Please correct me if I'm mistaken here

liuliu 4d ago

MoE, especially these Chinese ones trained on low-bits already is actually quite quantizable to lower bits, unlike smaller / dense models.

seanmcdirmid 4d ago

Also Expert MoE offload (to SSD), which still doesn’t let me (64GB) run a flagship but if you have 96GB you can.

liuliu 4d ago

The problem with SSD offload is the prefill time suffers, which is not really obvious.

seanmcdirmid 4d ago

I wouldn’t know since I haven’t found a model to test it on, but in my experience with smaller MoE models (35b A3b), they are the only choice for my M3 Max because prefill times are reasonable (I don’t even need to quant my KV cache, which would mess up MTP anyways). The only reason dense models (3.8 27b) are unusable for me are the high prefill times (mitigated if prefix caching is working correctly).

sheephess44 4d ago

MoE weights are much friendlier to low-bit quants than dense weights, because only a few experts are active per token so the per-expert error stays local. But the trap people fall into is that once weights are down to 2-3 bits, the KV cache and activation buffer at long context become the dominant term — further weight quantization stops moving the number. Memory-planning for MoE is really a context-length problem wearing a quantization costume.

tony_cannistra 5d ago

Not exactly your average "desk", but still neat.

  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.

bensyverson 5d ago

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.

mbreese 5d ago

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.

hirako2000 5d ago

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.

bensyverson 5d ago

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.

cyanydeez 5d ago

Apple and NVIDIA prices are not going to fall.

Their competitors will. Anyone locked into these two ecosystems are willingly going along with it.

bensyverson 5d ago

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.

Danox 4d ago

Apple’s problem is memory the biggest question is whether or not they will design and engineer around it their recent history says they will.

LoganDark 5d ago

There was the M3 Ultra 512GB Mac Studio, though.

mft_ 5d ago

…which is in the region of €20k on eBay, so the calculus doesn’t really shift…

LoganDark 5d ago

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.

unshavedyak 5d ago

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.

mft_ 4d ago

Apple’s article isn’t targeting people who bought multiple M3 Ultras before the RAMpocalypse - those folks can be safely assumed to already be on board with running large models locally.

bitcurious 5d ago

Puts into perspective how cheap the frontier models are via subs or even API.

mlsu 5d ago

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.

yndoendo 4d ago

It is the locally ran models that get me excited. They have an immutable user agreement and a focus on privacy.

Subscription / on-demand polices are a moving target, along with cost.

I foresee the ML market replicating the video streaming services market. Join us at a low rate and no advertisements. Six months latter, advertisements for everyone, unless you pay more $$. Now $$ will still show you ads. We are raising the rate to $$$ with adds ...

chaostheory 4d ago

The main subsidy, from OpenAI, is ending 10/26. It’s still “cheap” for now compared to all the other non-open weight frontier models, but it won’t last since OpenAI’s stated goal is to make it so that there’s no much of a difference price wise between on demand API calls and subscriptions

janderson215 5d ago

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.

Someone 5d ago

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.

antipaul 5d ago

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...

Danox 4d ago

Tim Cook, smiled. with OpenAI and Google a smile is easy when it isn’t costing you anything on the Apple side so John can smile away in addition, the AI models are getting smaller and more powerful by the day something else to smile about if you’re Apple.

browningstreet 5d ago

Still weird, and against brand, for them to lead with this particular model from that particular company. The groans are still earned.

Someone 5d ago

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”

browningstreet 5d ago

Second paragraph: "In this post, we show how to run the latest frontier-class model — Muse Spark Open, with nearly a trillion parameters..."

enos_feedler 5d ago

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.

bitpush 5d ago

> 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.

dwaite 3d ago

> 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.

You can stand up to government requests when it is legal to do so. Eventually you have to decide to abide by the law or leave the market.

macintux 5d ago

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.

origini 5d ago

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.

robotresearcher 5d ago

They did, and few people bought them, so they stopped.

myrandomcomment 5d ago

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.

Danox 4d ago

They discontinued the X-Serve because they had no control and they canned Nvidia and AMD in time. The last straw was Intel their no on the iPhone chip led to the creation of Apple Silicon.

They also had no’s from Motorola and IBM before Intel. All of those conflicts contributed to Apple creating Apple Silicon, which in turn led to in-house modems, and several in-house custom chips for other Apple devices down the road, I anticipate it will also lead to some move in the memory area.

dwaite 3d ago

Apple sells a Mac minis to companies though - to the point that they were supposedly asked to avoid redesigns when possible so that new models could slot into existing rack solutions.

mrboatshoepants 4d ago

That first bar chart showing between 15-36 tokens/second is for a 8,192 token context. Thats not nearly enough context to be useful these days

thoughtpeddler 3d ago

To be clear, at 0.75 words per token, that's about ~6,100 words (roughly ~12 pages), which I suppose could be useful, but as other commenters have pointed out, you'd need some kind of orchestrated swarm/hive that could smartly fit payload to within that envelope

infamouscow 4d ago

Agent swarms have figured out how to mostly work around context window limitations. There's no reason the same ideas can't be applied on a smaller scale, at least in principle.

boodleboodle 4d ago

Could you please paste some links about this that i can further read? Really curious

throwworhtthrow 5d ago

> 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.)

porridgeraisin 5d ago

I run my toaster for hours on end too. Love my sandwiches.

lemoncucumber 4d ago

Bad comparison, if I needed to make toast I'd use a Pentium 4 not an M5 Ultra

iAMkenough 4d ago

Once a month, I run my toaster oven the entire night to take advantage of off-peak energy rates.

aenis 4d ago

Imagine if you had 4 of them!

wktmeow 3d ago

It can offset the cost of a space heater in the winter!

comandillos 5d ago

A 800B model that performs way worse than other 300B ones. Well...

otterley 5d ago

What are you using to determine performance?

wolttam 5d ago

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

villish 4d ago

Muse is significantly faster and more token efficient. That is pretty impressive considering how much larger it is. There is a lot of potential even though this release won't shake up the standings.

wslh 4d ago

Am I right that investing in your own ML hardware only makes sense when you are dealing with private data and confidential sessions, right? (Not counting development or tinkering work).

dwaite 3d ago

If you think of it as API access vs temporary hosting running the same open model, it makes sense to run the hardware yourself whenever you are sure it can be fully utilized.

magnusokik 5d ago

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.

kingcauchy 5d ago

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

macinjosh 4d ago

Many thought Apple should be competing with the frontier labs while they see themselves as competing with Nvidia.

bigyabai 4d ago

Right now, they're competing with neither.

nico 5d ago

Very cool

Has anyone adapted the open Muse model to work like a decision model?

Sort of like Laya but based on muse spark

herunan 4d ago

So Apple is the NVIDIA of open weight in this case

bigyabai 4d ago

Apple is the AMD of open weight. No LLMs of their own, but a large appetite for partnerships.

_pdp_ 5d ago

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.

AndrewKemendo 5d ago

Anyone know why they went with Meta models?

dwaite 3d ago

They went with several open weight models. Presumably they did that because they wanted their research to be reproducible (once the 512 GB M5 Ultra actually hits the market). Using say a Gemini model they have access to only due to their partnership with Google would not be as useful for people to understand what Apple did.

bigyabai 4d ago

Apple has too many arbitrary, pointless grudges for Ternus to take them all seriously.

Google partnership for Apple Intelligence, Meta partnership for local models, and allegedly NVLink Fusion support for M8 GPUs. Maybe the White House told him that it's time to start burying hatchets. I don't get the impression that Apple's prior AI strategy under Cook made any real progress.

Danox 4d ago

Not pointless, all of that led to Apple Silicon, and you seem to be ignoring the fact that it’s not costing Apple any money. None of their hookups. This so-called cooperation is costing Apple a dime. But it is costing the other side billions of dollars.

Smile, the models are getting smaller and better, every generation of Apple Silicon is getting better in the opposite direction. The only problem is the cost of memory and that is within the design engineering ability in-house it is also within their financial capacity, so if I’m John or Tim, I too would wake up with a smile because I can fix the one thing that’s giving us trouble memory because we have the talent and the money to fix it in the next four years.

bigyabai 4d ago

Apple Silicon copied AMD's mistake. RDNA and Apple Silicon are both the slowest GPU architecture for AI because they spent 6 years optimizing raster performance over compute. Now there's 4 generations of Apple Silicon Macs that rely on compute shaders for any kind of efficient inference. Even the M5 GPU is less programmable than Nvidia's ARM GPUs from 2017.

ARM datacenters were a trillion-dollar market that Apple made ~$0 off of. The Mac Pro could have fit that niche, but Tim Cook chose to depreciate it instead of letting the PCIe connectors be used with Nvidia cards. He cut off Apple's nose to spite Nvidia's face. It's nonsense, especially since they're crawling back to Nvidia to license their interconnect, in tacit recognition that it wasn't a real grudge in the first place. My guess is that Ternus doesn't have Tim's money allergy, maybe at the White House's request. In any case, the grudge is over, and Apple forfeit ARM HPC altogether.

> because I can fix the one thing that’s giving us trouble memory because we have the talent and the money to fix it in the next four years.

By all means, continue. Which X account or YouTube channel convinced you that Apple could replace SK Hynix in 4 years?

dwaite 3d ago

> The Mac Pro could have fit that niche, but Tim Cook chose to depreciate it instead of letting the PCIe connectors be used with Nvidia cards. He cut off Apple's nose to spite Nvidia's face.

Apple Silicon only has support for 16k pages via DART, rather than the 4k pages typically used by AMD/Nvidia. The requirement to coordinate a major change of drivers to understand 16k pages is why both PCIe GPU and eGPU do not work (or rather, did not work - TinyGPU is a compatibility layer that adds eGPU support and is used for AI workflows).

The Mac Pro's biggest redeeming feature for a data center was that it was rack-mountable. The ARM version was still mostly an empty case. It remained mostly for its ability to support PCIe cards for video and audio production use cases, such as live broadcasting.

> especially since they're crawling back to Nvidia to license their interconnect

Licensing the interconnect for broader compatibility. They support other interconnects as well.

pjmlp 4d ago

Yes, the traditional Mac Pro customer now has pivoted to Windows or some Linux powered enterprise workstation powered by Ubuntu, Red-Hat, SuSE.

AndrewKemendo 4d ago

That makes sense I also assume there’s better allyship with Meta than X long term geopolitically and they are already in bed with google on search

openai and anthropic are too chaotic to bet on would be my guess

isodev 5d ago

Preparing to upsell those very expensive Macs coming later this month, are we.

butterlesstoast 5d ago

I -- love -- an article -- full of -- em -- dashes.

shagie 5d ago

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.

NekkoDroid 4d ago

Do note that those are en-dashes and not em-dashes.

ManuelKiessling 5d ago

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!

jonathaneunice 5d ago

Em dashes—love ’em!—but they do. not. need. personal. space.

Apple ML Research: Muse Spark Open with MLX on Mac Studio (machinelearning.apple.com)