I really do not understand the value of running large models locally. I am absolutely fond of local models but for specific tasks. The agents need to do a lot of work in this case: orchestration, big or small model split of planning and work, split into small tasks, merging, etc.<p>LLMs with <30B parameter models seem like the sweet spot. I would rather focus on multiple tiny LLMs that are tuned to specific tasks (SQL queries, finance/accounting, backend/frontend dev, etc.).<p>For larger models, a $10-40 ish Neuralwatt/Opencode Go or any similar subscription goes so far for the average user that it is ridiculous. We are still in the "youth" era of agents. There are many patterns that work well with smaller LLMs. Or with a big-small LLM combination.<p>My context:<p>I own the tiny 16GB M4 Mac Mini and I use LLMs (coding agents) to build software and a few of my products use LLMs directly. Full-time user of LLMs, do not read or write code manually. I run the tiny Qwen series (<10B params) sometimes but a couple $20 subscriptions (Claude, Neuralwatt) take care of my full-time agent usage.
Apple, to me, is the winner in the end of this AI chapter.<p>Distillation works: we can get models leaner, have better performance with lesser and lesser parameters, more efficient, running locally. Apple gets this and we'll see the result with M7 Max/ultra chips and I reckon this will be one of the titular apps when M7 line of chips are introduced.<p>I guess it explains the haste Anthropic and OpenAI are in to IPO, cash out and leave the the retail investor holding the bag.
I think the question to me is whether Apple can ship affordable consumer devices with enough memory to run sufficiently useful models sufficiently fast.<p>Right now 128GB MacBook is the top of the line and you could do Qwen 3.8 Flash or similar.<p>There is a rumor that the M7 chip can use 1.5TB of memory which would imply Apple is aiming for more memory on devices.[1]<p>Would that imply the entry-level MacBook will have 128GB and top of the line models could run larger reasoning models? Then the question is price and if any of this will translate to iPhone for on-phone models.<p>With how expensive memory is, IDK if they can pull this off but I hope they do<p>[1] <a href="https://www.tomshardware.com/tech-industry/semiconductors/apples-rumored-m7-ultra-targets-1-5tb-of-memory-and-blackwell-class-ai" rel="nofollow">https://www.tomshardware.com/tech-industry/semiconductors/ap...</a>
> Would that imply the entry-level MacBook will have 128GB<p>Most people neither need nor want to run large LLMs locally.<p>The entry level MacBook Neo costs $599. Even if RAM prices revert to the pre-AI mean, the entry level models are not going to be that stacked with RAM.<p>Eventually we'll get to a doubling where entry level means 16GB of RAM, but most people won't be running anything more than simple models for light local work.<p>Have you ever tried running large LLMs on a MacBook Pro? It's hot, loud, and makes your battery level disappear at rates you didn't know were possible if you're not plugged in. It's a very niche hobby.
> whether Apple can ship affordable consumer devices with enough memory to run sufficiently useful models sufficiently fast.<p>i suspect that they will at least do offloading for casual usage as well. if they can ship 24GB in a homeview or appletv style product, it's game over.
Even if it's 2030 - that's a mere 3 years from now - it will take at least that long for the greater economy to figure out how to use LLMs. And probably a memory boom/bust cycle in that time too.<p>If Apple, in 2030 has a compute product that a worker can BYO(X) to work that has 2TB RAM, with an encrypted personalized context that can float between agents, then they will be in a very interesting place.
The Mac Studio hardware can do 512GB. That might be a hardware limit of the current architecture.<p>But the family shouldn’t be limited by that. We’re not going to see M6 Max hardware, because the M7 line is being fast-tracked and pulled forward. And I wouldn’t be at all surprised to see M7 capable of 1-2TB or more.
> That might be a hardware limit of the current architecture.<p>It is, at the very least, a practical limit. There are unlikely to be enough customers willing to pay $50k+ for a hypothetical 1TB RAM version to make it worth manufacturing.
The limit is strictly to cost of memory at this time. If you wanna get by that limit, you have to, well… design it yourself in house.
Do you <i>really</i> think that Micron, Samsung and SK hynix are working in a cartel to prevent DDR5 memory from being manufactured?<p>Spoiler alert: they're supply-constrained too. Their customers will pay to outbid Apple, just like how Apple outbid the industry on TSMC's 5nm node.
M5 Ultra has 64 channels and the densest LPDDR5X chip is 16GB - so they could produce 1TB Macs but I doubt Micron or Samsung has any to sell.
Why is there a desire to have this stuff on the edge on the engineers lap instead of just building some arbitrarily scalable cluster for I'm guessing cheaper than apple's usual premium? Habitual practice from the engineers who would compile on the device getting hot in their lap? Lack of widespread hpc experience compared to running docker?
I don’t want it in my laptop, I want it in my mini server rack at home. I’m not running any of my personal data through an external service that is certainly collecting it as training data.
Many users would prefer to use a Sonnet-level (or better) LLM without sharing their data with others.<p>And with a device like this, you pay one single time to buy the product and then never have to think about paying for tokens ever again. For ordinary users that's silly, but for users or businesses chewing through lots of tokens, this means they can create their own turnkey replacement for "the cloud" where the only cost after purchasing the machine is the electricity to run it (which with these machines is very little). Mix that with the privacy benefits and suddenly it makes a lot of sense.
Seems everyone in this thread is assuming I am saying go to private third party compute. No. Build your own cluster on metal you own on property you own, and have your team use that and just some thin clients to write with. Your compute goes way further. Your cluster can be going 100% of the time with jobs queued up by your team. All that memory you buy for your team in macbooks is idle and unavailable when your employee closes the lid and tosses it into the bag into the car on the way home. It could be used 24/7. On hardware you own. For cheaper to build I'm sure than apple hardware tax. On easier OS to manage and no doubt deal with anything involving any sort of heavy compute work than OS X. Are we really going to waste resources with docker when we could be using a node completely? Come on.<p>Like I said, seems HPC is lost art. All four sub replies missed that in my original comment.
yes. People do not read.<p>I agree with you, and i would like to set up a server at work for local inference but we simply do not have the knowledge, it's much easier to get a couple of AI Max machines that are good enough for me and the boss which i convinced on the good points on running clankers locally and be done with it (though my plan is to use the AI MAX machines to figure out what the infrastructure and workflow will look like, and at some point ask for the investment)
People don’t want to send all their private company data to the model providers?
Because why not? If you can get unlimited inference with roughly Opus 5.5 intelligence for $10k that easily beats current API pricing. Like not even a question.
Distrust in allowing an unreliable vendor to control the fate of your business?
Apple’s position down the road. When the Chinese company CMXT takes a much bigger share of the memory market worldwide, and they will. Their dominance in memory will probably be pretty close to what BYD is today in the worldwide car industry.<p>I don’t think Apple can be in a position where all of the Chinese device-making companies can get memory at a significantly cheaper price from CMXT, and Apple is stuck with buying memory from the three hyenas of memory at a significantly higher price.<p>That’s not going to help Apple sell their devices at a competitive price worldwide over time against the Chinese computer makers (Lenovo, Asus, and Motorola) and make no mistake about it: their share of the market worldwide will get bigger because of this change in the memory market; their focus will be on the long term, while the lawyers/MBA/finance types in the West will have their heads stuck in the short term.<p>It’s the (Chinese companies’) opportunity, and they’re not going to let it pass; I wouldn’t. In addition, Apple is in competition with Nvidia, Microsoft, Intel, AMD, Qualcomm, Google, Meta, and indirectly with Anthropic and OpenAI (whose CEO, Sam Altman, recently declared that computer ram is wasted on low-level regular computer users), along with a whole legion of other 3D model makers on the procurement of memory.<p>The combination of all these elements going into the future means that if Apple wants to stay ahead, I believe Apple has no choice but to take memory in house, they have the design and engineering capability in-house, and they have the money because they didn’t burn it on AI models and data centers.<p>This scenario is no different than replacing Intel or Qualcomm; you can see the writing/destination and the ultimate train wreck up ahead if Apple doesn’t roll up its sleeves and get busy, the market has changed and again Apple has to change and adapt to the new conditions.
I'm on 128gb Mac and Laguna2.1 is able to do 800tps coding as good at Sonnet 4.5 or Opus 4. It's awesome.
Most likely Apple has been spending whatever they can to get in front of the memory market.
They <i>definitely</i> are, judging by their price increases.<p>Even so, DDR5 is so high-demand that it's likely that Apple can't outbid customers like Nvidia or OpenAI when they drive the margins they do.
I dunno, I reckon they are targeting 64GB and will develop a smaller model thats built in by default that handles a lot of simple tasks quickly.<p>Probably sitting under a router, so that it can hand off bigger jobs to an Apple DC somewhere, probably with a certain number of credits based on your iCloud sub.
They'll win if they vertically integrate chip manufacturing - particularly the memory.<p>It's not as if they are short of money.
Isn’t Apple already doing this? Or what kind of vertical integration do you see other than their unified memory baked directly into the CPU?<p>Or are you saying that Apple should become a memory manufacturer, rather than working with TSMC?
Apple does not manufacture its products.
> I guess it explains the haste Anthropic and OpenAI are in to IPO, cash out and leave the the retail investor holding the bag.<p>I don't think so. LLM inference enjoys huge economies of scale. At no point in the foreseeable future will a local model be able to compete with a cloud model in terms of price. Local models will have the privacy angle, but very few people will be willing to pay top dollar for that. There will be some people, for sure, but the vast majority will just use cloud models.
Why do you need huge economies of scale though? For 90%? of queries a local model will be good enough?
Well beside for roleplaying, local models are a bit useless, especially the ones running on end user class hardware.<p>Maybe when chip would infer at 10-20x the speed of a 5090 RTX, otherwise it's really..bad.
For inference alone they get far better utilization of the hardware because they can keep it fully loaded with maximum concurrency. Other than that they get to collect a lot of data for training.
Normie queries can be handled quite well by Haiku 4.6 or whatever.<p>What new bike should I get? How should I rephrase this email? What to do about my persistent headaches? Recipes based on these ingredients?
Small models inherently have less pre-trained knowledge, which will hurt their performance on queries like this. Moreover a lot of queries benefit from doing web searches, especially if they need to respond to changing conditions such as your bike example. Now you have to send data through a data center anyway for good performance, and probably need to use an API if you don’t want your computer blocked. So what’s the point, when GPT Luna exists if you really want to get a decent model at a rock-bottom price?
Most people just Google something or use the free ChatGPT. They have no clue how the answers get there and frankly would be just as pleased with GPT 4o as they would be with Astra. Questions like “what does this light mean on my dashboard” or “which bidet is best”.
My guess is it will probably be both. There was a time when no one could envision computing without the mainframe.<p>In the mainframe's heyday, it was difficult to grasp a world in which compute became so dense and so cheap that it made sense for each individual to have all processing and memory on their desk, talk less of on their lap and in their hand.<p>Of course this time could be different. But, when we look at previous relentless progress in compute and efficiency, then also couple that with the breakthroughs AI itself unlocks, it's hard for me to imagine that centralized AI will meet both the scale and flexibility desired for all use cases.<p>Just as centralized compute never fully disappeared -- it took the form of client-server computing and the cloud, there's no question that centralized inference will continue to exist.<p>But it wouldn't surprise me if, ten years from now, 90% of global inference happens at the client.<p>While simultaneously, the 10% of inference that takes place in the datacenter is wildly lucrative and orders of magnitude greater than today.<p>Put simply: the inference pie may become so large that both client and datacenter win.
Unless cloud PCs take off and kill "AI PCs" and the future of the end-user hardware market, I think it'll be a bit of both.<p>It'll become easier and more affordable to distill/train/optimize smaller local models to meet your specific work needs. Multi-agent efforts will orchestrate the balance between using on-prem hardware for most tasks, and then defer to cloud models for more demanding/edge-case tasks. Especially if an organization can manage what data is shared with which model, and most of their user base are everyday casual users.<p>I also expect cloud providers to eventually start charging different rates for on-peak and off-peak usage, which would incentivize running what you can with the local compute you have otherwise sitting idle across dozens to hundreds to thousands of machines in your fleet.<p>Anthropic just changed their policy to invalidate any unused credits without refund if you buy your API credits annually. I see that as a sign of unpredictable budgets for cloud models for customers.<p>I could be wrong and all your future hardware purchases for end-users will be thin clients, without the local processing capabilities of today's laptops and towers. I don't see the hardware manufacturers profiting much from that, but maybe we'll see more companies like AMD declaring they prioritize cloud compute customers over everything else.<p>With a hardware refresh for cloud computing every 3-5 years, we'll likely see today's compute hardware driving Opus and Sol available on the secondary market for companies willing to build more on-prem capabilities.
Cloud compute has been completely commoditized. Yes, local models will get better and better, but I think for the vast majority it will continue to be both <i>cheaper</i> and <i>more productive</i> to use cloud based models, far into the future, where bulk purchasing, oversubscription, and discounted industrial electric rates can be taken advantage of.<p>Where I am, it's cheaper to purchase <i>multiple</i> claude subscriptions than it is <i>for the power bill from running a local model</i>. Factor in the hardware costs, and you're looking at a break over a decade in the future, for comparable cloud models. Yes, there's privacy, etc, but now you're down into fractions of a percent of users who care about the government reading the memory of their computer.
In the end, I think companies like OpenAI and Anthropic will be the nameless and faceless backend provider, and they will get decoupled from the users. At which point, if they haven’t already had their multi-trillion dollar IPO and everyone cashed out, I don’t think they’re going to.<p>Apple and Microsoft and other companies will own that front-end experience with the customer, and it’s not going to matter to most customers whose backend gets used for what.
I agree with you, even if many of the pundits say that Apple was left in the dust.<p>At some point, the real cost of inference/intelligence is going to come due, and the existing open source models will be good enough for 95% of tasks users want (or training their own model(s) will be inexpensive for Apple).<p>Apple will be there, as usual, to provide secure, on-device, faster inference that the vast majority of their customers will be fine with. Let everyone incur the R&D costs.
This is a nerd fever dream. Yeah I’m sure people can’t wait to run local models on their iPhones.
For those of using AI day to day, not being beholden to the whims of Anthropic and OpenAI is definitely a dream scenario.
Is Apple Intelligence not local models currently running on iPhones?
You don't have to sell it as a local model. Sell it as offline capabilities. Sell it as private background processing while your phone is idle and plugged in. Sell it as un-enshittified LLM once Anthropic and OpenAI inevitably screw everybody.
Take a look at how qwen3.8 27B compares on benchmarks to much, much larger models. A while back I called it the crow of AI models: paradoxically strong given its tiny size, like the brains of crows and other intelligent birds.<p>It's not comparable to huge models of course, but given the power/size ratio here it shows that serious gains are on the table for models that can run fast on a high-end laptop.
Man, if all that makes Apple a "winner" then I'm dying to hear your assessment of Nvidia.<p>Factoring in the increased cost of memory, Apple's threatened service revenue and the stagnation of ecosystem products like the Mac and iPad, I think they have a really challenging future to navigate. The whole "sell people a B2C ecosystem" business model is more threatened than ever.
The design and engineering of memory is not more difficult than replacing Intel and Qualcomm, Apple certainly has the money, and they certainly have the engineering talent within the Apple Silicon design group. It took 13 years to replace Intel and six years to replace Qualcomm who has already warned their investors that they anticipate losing revenue in the near future. The question is whether or not they have the will to do so, money and engineering ability isn’t the problem, it is simply a matter of will. With CMXT coming online bigger and stronger.<p>I don’t think there will be a choice long-term for Apple the conditions of the market have changed, and you have to change with it or get left by the wayside. They have shown that they will adapt and they certainly haven’t wasted money like their tech competition over the years.
Difference being, Apple had experience designing CPU cores and wireless radios before ditching Intel and Qualcomm. The transition felt "magical" because it was a generally low-effort layer on top of a muscle that Apple had gradually built over time. Their design approach is not compatible with scaled manufacturing; Apple does not have experience manufacturing DRAM at all. They don't have their own fabs, nor could they bid on fab space when AI companies outspend them 10:1.<p>Apple would make more money investing in a CUDA-style architecture than they would by becoming a subsistance manufacturer for the pitiful margins of the iPhone, iPad and Mac.
The contrast I'm drawing is running these models locally vs on the cloud and how that would be a selling point well received.
You are crazy. Apple has no lift on AI local or otherwise. Local LLMs are a massive and stupid waste of yes YOUR time.<p>And Apple really wants to sell people who have no cred in the space their stuff because THEY have no cred in the space.<p>Its basically a club for stupid f**s
Sure, all you have to do is use only what can climb the walled garden
[flagged]
"Harbor is currently in a private preview under Apple’s Claris brand"<p>There's a name I haven't heard in a long time! Claris (always an Apple subsidiary, since being founded in 1987) renamed themselves to FileMaker in 1998, but apparently they rebranded back to Claris in 2019.<p>Looks like they've mostly just released annual upgrades to FileMaker for the past decade, but they did add MCP to it last year: <a href="https://www.claris.com/blog/2025/why-mcp-matters-for-claris-filemaker-developers" rel="nofollow">https://www.claris.com/blog/2025/why-mcp-matters-for-claris-...</a>
> Cluster for scale: “Connect two or more Thunderbolt 5-capable Mac computers and Harbor pools their unified memory into a single inference resource. Run models that exceed any single computer’s capacity, or serve more concurrent requests across your cluster.”<p>This seems like the main way it would differentiate itself over existing solutions like LMStudio etc.
This is what I've wanted for the last few years. I have played around with local models and had some success, but I don't want to deal with the hassle of setting up, updating, etc.<p>If all of this were easy, I would happily invest in new dedicated hardware to run it. This could be a full-blown Mac or it could be a souped up HomePod-type device, loaded with RAM. Better yet, make it a revitalized Matter-enabled Airport!
MacMini with M7 Pro and 96G of RAM.<p>They are going to sell sure 10,000$ and you won’t need anything else for the rest of your life (if you are over 90).<p>Btw, llama.cpp with llama-swap ain’t that hard to configure if you have access to a potential LLM and have 5h of your life available to that task.
I love that they are bringing Claris back for this. Moof!
I really want capable models local but it’s impossible for me to afford the hardware. If there is a magical way to run glm 5.3 without quantization at 80-100tps with a total spend of less than 10K I’d be happy but that’s impossible now.
$200 a month is only $2400 a year.
That’s 4.16 years until you spend 10k.<p>I don’t find it reasonable at all.
Subsidized Subscriptions or api for cheaper models, is the best value.<p>But the future of AI is so uncertain, regarding model sizes. I don’t think making huge hardware investments is reasonable for individuals.
You don’t get privacy at that price.<p>There is no ZDR at that price.<p>I would not let a non-local AI have access to my financial or health data.
> $200 a month is only $2400 a year. That’s 4.16 years until you spend 10k.
I don’t find it reasonable at all. Subsidized Subscriptions or api for cheaper models, is the best value.<p>If you buy your own hardware, then after those four years you still have that hardware, and it likely still holds significant value. Plus you know the local model you're running today won't change or suddenly degrade in four months, let alone four years. Whereas after four years of paying subscriptions and API costs, you have literally nothing.
After I think about it, this makes sense that Claris would tackle such a project under the auspices of Apple. They're a subsidiary so they can monkey around with this stuff without having to adhere to the draconian everyday operating restrictions of "Apple proper". Good on them if they can pivot into something that hews to AI.
"Harbor is currently in a private preview under Apple’s Claris brand, but it seems likely that we’ll see it launch to the public in the weeks and months ahead."<p>Claris?? Ok, this sounds like it's just someone's personal project. Something like this that was actually going to ship would be in a different org.
<a href="https://getlocalcode.app/" rel="nofollow">https://getlocalcode.app/</a> runs local AI models (larger ones of course, DeepSeek 4.1 Flash 1017) easy. You can download and try it now. The problem is elsewhere: you cannot find M5 Ultra that you can purchase and set it up today.
You can chat with the local model that ships with macos on your M1+ mac right now. There's a few servers on github that make it easy.
You can also "fm chat" in terminal
Could you elaborate please?
Can you shut that thing off?
(1) All the headaches of maintaining FileMaker Server on Apple hardware, (2) multiplied with the additional requirement to cluster devices, (3) but without users accustomed to the silliness of misusing consumer devices running a consumer operating system for a service that must be reliable, and (4) an Apple subsidiary unable to bring about meaningful change.
An Apple M5 Ultra running an app like that? Because Apple hardware is a monolith, unlike the PC space, "Harbor" will be configured out of the box with the ideal settings for many well-known open-weight models.<p>... Looks like I picked the wrong week to buy parts for a linux-based inference box.
OMG after 25 years, Claris gets something beyond FileMaker (among the most legacy software still alive)? Claris would not have the engineering talent, so why would Apple use them as a vector?
What would make running local models easy is being able to buy a Mac Studio without a 6-month delay :-(
But you’re still in for a ton of pain if you want to run them headless right?<p>They desperately need to fix that.
It was pretty fast to get it set up. It could be better for sure, as it's not a first class concept the same way that it is in linux, but an hour of google search ai mode will get you where you need to go.
I know you can set it up to make it seem to work, but from what I’ve seen people say over and over, it will eventually end up asking you some question through a dialog box. And you will have no idea that’s going on other than all work stops. And the machine seems locked up or something until you connect a display and find out it wants you to click OK.<p>It SHOULD be totally possible. But it fights you.
How so?
MacOS is a desktop OS. You really REALLY need to treat it like one.<p>Lots of things pop up and can’t be dismissed any other way than though mouse/keyboard on the display.<p>OS updated and wants to tell you it worked? No command line way to get rid of that. Program wants to do something and one of those system security prompts pops up (“BBEdit wants to access your Documents folder”)? Same problem.<p>You can set it to auto login and auto update and grant permissions and everything.<p>And then, one day, it will stop responding and you’ll have to hook it up and log in to see some pointless dialog that just wanted someone to click “OK”. And until you do no work gets done.
Deepseek has found safe harbor!
Why so much investment in local models when the remote ones are so cheap ? Everyone was fine with remote compute & storage before AI, now they balk?
[dead]
Hopefully it won't be another preinstalled bloatware app like Image Playground. macOS enshittification has been greatly accelerating, even giving Win 11 Home a run for its money.
[flagged]