It's a niche market but it's a market that overlaps heavily with professionals in the AI space and lead developers, so it's a market that gets them customers in those roles.
If I were running Apple I'd call the RAM price bubble for what it is and temporarily eat some margin to offer machines with more RAM than competitors, especially these models that are great for edge AI, and capture market share.
Nvidia has CUDA, AMD has CDNA, and Apple has... compute shaders, I guess?
How much it matters in inference? Most GPUs have enough computing for that and the bottleneck is the RAM speed and size. And M5 Ultra is becoming to challenge this.
Is that what others read ?
This unified architecture makes Apple hardware very good for AI work, where latency between RAM and CPU is very important
Even the SSDs (which are normally more strippable) use a proprietary hardware form factor.
I think the article mentioned the real draw:
> Apple noticeably promoted the ability to link multiple Mac Studios together into a single, more capable system for running large frontier AI models, a feature aimed at business and developer customers rather than everyday consumers.
(Couldn't read the source cited, might have more info: The Information)
It's kinda why memory bandwidth is an enormous red herring, even for datacenter applications. Nvidia's huge advantage is a compute-optimized GPU architecture and their Infiniband networking, their memory controllers aren't really the star of the show.
I'm training a model using reinforcement learning with self-play. I can and do use vast.ai when scaling but for experiments it's far faster, and cheaper, to run it locally until the bugs are all figured out. Just provisioning a new instance and copying the relevant checkpoints and things can take 25 minutes. It's zero locally.
Inference with MLX is surprisingly zippy. I’m running a classification task on the entire HN comment dataset and it’s projected to take about two and a half days, which is not bad considering we’re talking about tens of millions of comments.
Yes, I could do it much more quickly by throwing Modal GPUs at it but this is low-priority work. I might as well throw my M4 a bone.
Apples ProRes codec is only licensed to run in high quality mode on a Mac, and so my Nvidia PC can’t do what I need. Thus, I own the beefiest Mac Studio you can currently buy. I would pay more for more TFlops.
I have done local LLM on there but it wasn’t interesting. Far worse performance and intelligence per dollar than the cloud boys.
There is no cloud offering for my video needs though.
If you can fit it on a GPU, and especially for training, it is so much quicker than a Mac.
I realize I’m somewhat limited (16GB RX 9070), but still, it seems really far off from the kind of experience even a basic $20/month subscription gets me.
Any tips anyone might have are appreciated! I’d love to be local first and would be willing to buy hardware to get there.
I have an older M2 Mac mini that does the OCR and visual description of all my screenshots. Screenshots are stored on my NAS.
I like to screenshot things as a quick way to remember. They are things that I would not be comfortable sending a cloud provider (customer data, prototype screenshots, bank dispute details).
It runs Qwen3.5:9b and glm5.2-ocr with Ollama and uses about 10GB of RAM. It automatically releases the models from RAM after 5 minutes of inactivity so it is pretty seamless to leave running in the background.
All the details are stored in a simple webapp with a SQLite db that I can search through.
Doesn't Apple do this already within it's OS all locally? It certainly does it for OCR and categorization.
EDIT: Also, no reason to use a generic LLM for this. This functionality exists in something like Immich (both OCR and 'context categorization'), and doesn't tie you into the Apple ecosystem either.
The $20/month subs are much stronger than the local models you can run, even with how far local models have advanced lately.
The appeal of local models is that the data never leaves your network so you can feel safer putting sensitive content into it. It also feels “free” to use when you’ve already paid for the hardware.
But it doesn’t perform better and if you do the math you’re probably not saving money either. It’s helpful for things that you can’t or don’t want to outsource to a 3rd party.
- translations: cloud providers can bowdlerize (censor) bad words/content; also, if you want to do a translation for personal use of copyrighted materials, cloud providers may block it
- image generation: generating drawings with a style that even just resembles a copyrighted one (ie. Disney) may be blocked by cloud providers - for example, generating old cartoons style with GPT may not be possible.
but depending on what you're doing, you may not need the "bleeding edge" performance
All of this is private, but not necessarily sensitive. You never know what is happening with this data. They might say they don't log it or don't sell it, then few years later you'll find it all online or read a book that has a story eerily similar to what you chatted about with GPT a year ago.
IDK if that might be a concern for Apple or their AI partners.
I just ordered a new Mac Studio M5 Max 128GB $5899 ($6400 with tax) to be able to run the bigger "consumer size" models in the 70B parameter range (~96 GB). That said, I have no illusions that this expensive setup with a Qwen Flash coding LLM will be comparable to a $20/month subscription. Even upgrading to an even more expensive Mac Ultra 256GB for $10000 to hold a bigger model still won't be comparable. Apple hasn't shipped my Mac yet and I'm still considering cancelling it and downgrading to a smaller 64GB RAM config ($4299) to save $1600.
Why did I initially spend the extra $1600 if I knew ahead of time that it wasn't as good as cloud AI? Because I thought I could use some local LLM for the easy tasks or when I hit cloud rate limits. No issues with privacy so that wasn't part of the motivation at all. I just wanted some local AI capability to augment a subscription. I've not totally convinced myself of the cost/benefit of this.
Based on today's consumer hardware landscape, you're paying very high prices for crippled capability compared to the cloud AI subscriptions. We're also in a transition period where the next iteration of hardware improvements have some compelling features for local AI. Apple's upcoming M7 (2027 or 2028) is anticipated to have better GPU and neural engine to help with prefill TTFT. AMD Strix Halo is about to release 192GB system which is a big upgrade to their current 128GB ai pc. Maybe apply my $1600 savings towards those newer products. Those future products will still be very expensive but maybe the cost/benefit will be better.
The maths don't check. With Deepseek Flash one goes a very long way with 1600$ - even 10$/month, for easy jobs, are more than 13 years, and at a higher quality.
Big upgrade to memory capacity but memory speed is only going up by a few percent, so its still going to be slow with more than a few B active params (I have one)
- There's no guarantee of the $20/month service, and it likely has some limits compared to dedicated hardware token wise.
- Model are becoming more and more efficient, in many cases an M1 Max Mac Studio is still capable with 32 GB. 128 GB ram may not be the necessary baseline.
- Folks may think they want to only have a general model running locally (it's the comparable after all from the cloud providers), but we have to remember if the tasks we're trying to do ultimately are more specific than general and if there's space for the smaller models to do that.
LLMs are harder: not much useful below 12B, and the 700B+ ones are really much better. Models like Qwen 3.8 27b show promise: in a few years pretty good local AI should be in reach for anyone willing to buy a $1000 computer (but who knows what your $20 sub buys you then).
We also built some QA agents that are always playing our games from the same builds a player would and flagging things to fix/improve; that alone needs the game focused and front-and-center so it can properly screen-capture for deciding what inputs to take next (and for screenshots/replays), which also means we can't really do any hands-on work at all on the machine when it's running.
Having a separate (and tiny) machine for all of this has been great. We don't bother with local models because, you're right, the $20/month sub is way better than anything that can run on small consumer hardware atm.
I'm curious about your setup. I've been tinkering with the idea of setting up Blender (cli use) in a container to allow agents to verify the scripts they are generating compile at a minimum. One thing I've found extremely helpful was generating a RAG of the current version of Blender.
For anyone wondering, I'm running Gemma4 26b A4B on a mini PC with 32 GB of DD4 and a Vega 7 iGPU (llama.cpp w/ Vulkan).
I realistically costs $5-10k to replicate a ChatGPT like agent. And it doesn't scale.
That's still really close. And models and quantization etc keep improving.
I'm absolutely positive that I'll be switching to mostly local AI in the next 5 years.
If you have a real product and can actually sell it, youre taking a largish risk relying on the cloud.
From model changes, alignment, to enshittification and the natural cognitive offloading, you could be one day removed and ROI tanked.
Think of AI like a mafia boss who helpfully supports you untill they need a favor. Thats all cloud AI is in America.
I have been eyeing a 512 GB Mac 5 Ultra to run full DS4 pro locally, which I expect would be pretty amazing as far as quality/recall. The only downside is that the speed is a lot slower than something like 27B on the 5090.
What I noticed is that (1) the great local models are optimized run inference (diffusion & LLMs) well on 32GB VRAM <= GPU's because that that's what the target has ...
(2) The quality of local models (esp. in diffusion) is increasing faster than the need for more VRAM - additional reason for the value of these FAST GPUs to increase!
(3) RTX PRO 6000 96GB is really great for fine tunes (ai-toolkit) :) but doesn't outperform my RTX 5090 with inference by anything significant on the good local models.
I have never run an AI job on a Mac, i also have doubts about performance and compatibilities - since the reviews almost never compare directly.
The model I’m specifically targeting to use at high speeds is Qwen 3.8 27b @q4ks. This model actually proved to be good at coding (it sits somewhere between Sonnet 5 and Opus 5 capability). M1 got 10 tok/s, Ryzen Halo 20tok/s, and Radeon 7900 XTX 50tok/s (can only do 128k context window in Radeon card).
The prefill gets extremely slow around 50k tokens in context window (whatever prompt processing stage entails could be wrong about phases here). It takes about 2 hours to fill the context.
Even with a drafter model intended for speed instead of mtp, I can’t get past 70tok/s, still is extremely slow to process prompts as context grows, and drops down to 40-50tok/s anyway making this config still moot for improvement on my Radeon card.
The only thing I can point to slowing me down is bandwidth of the card itself.
I am waiting to actually get my 5090 right now and I am betting that the 1700 Gbps of capacity will fix my prompt processing speeds. I don’t need full PCIe lane bandwidth to serve my house I just need to load the full model into vRAM and let the GPU do its thing.
Additional benefit to the external enclosure route is being able to migrate the inference between devices more easily. I can develop out the infrastructure then migrate the card to be hooked up to a shared node in the house with all the tools necessary for my family to take advantage of the privacy enhancement that comes with local inference.
So if I stay within 35B, especially MOE, my M5 Pro 64GB MBP can also run them well, and it can do plenty of other stuff too including gaming. While 256 GB with such RAM bandwidth and powerful GPU sounds like fun on paper, it doesn’t seem to be the next level compared to 64 GB
Really curious what people run on 256 GB Macs
if they are a heavy user, perhaps they string 4x together.
I've been running into annoying limits with Claude recently. It gives me like 5 questions over the course of 15 mins and then tells me to wait 5 hours. When companies can change things up to make the base subscription nearly useless (the last question always gets messed up, too), then you realize the value of owning your own infrastructure.
I've spent $2 in the last 2 weeks on OpenRouter. I've been trying to only use the medium sized models that I would otherwise be able to run on a nice local setup. That nice local setup would cost ~$4k. I don't know what the operating cost would be, but I would be concerned that my home electricity would cost more than at a datacenter. It just doesn't make sense right now except for privacy reasons.
I'm probably going to hoarde open weights models in the ~31B range until memory costs fall in a few years. Then, I'll buy some hardware to run at home just so I feel more sovereign over my stack regardless the cost/token speed.
But I am looking forward to lower hardware costs!
For OpenAI and Anthropic, the $100 subscriptions cost 5x the $20 subscriptions and give you 5x the tokens. And the $200 subscriptions are 10x the cost for 20x the tokens. (Tokens cost 50% as much.)
I recently built a minimal Dark Software Factory out of an N150 Mini PC. It uses three models; Sonnit, Sol, and Gemma.
But, I have a LOT of instructions about how I prefer the software it builds. Gemma doesn’t handle all my instructions very well. But it’s close!
I’m running gemma-4-12b because I have limited RAM and larger models were too slow.
I do two types of jobs: planning and prototyping. It has done fine at some of my planning rounds.
I still consider it experimental and don’t use it a lot but I think we’re getting there.
Though keep in mind not being beholden to shenanigans from said cloud companies (and interference from government entities!) is definitely worth something intangible.
Local llms don't suffer from cloud availability issues. Anyone that used Google models know that sometimes they just don't have capacity whatsoever, at least that was the state of things some months back when I used them. Just bear in mind if needed, cloud providers will prioritise API and corporate customers over subscriptions if availability degrades more.
Also they don't have the same guardrails as the other models, so for hacking, reverse engineering and black coding (piracy etc...) these local models might be the only options.
But currently it's really hard to beat anything offered by the cloud companies. And the cost and complexity of setting it all up, just to barely (if at all) touch on Opus-level intelligence makes it seem like we're not quite there for the common man (enthusiasts are a different story.)
I am very excited for open source local models, and we're nearly there, but it's still too complex and expensive to be my daily driver (yet).
My understanding would be that if you’re interested in this sort of card for AI that you should go with the AI PRO R9700, which is basically the professional version of the RX 9070XT but with 32GB of memory.
It’s significantly more money but not crazy like a 5090.
I just happen to have the 9070XT primarily for gaming purposes.
I’m not quite sure how to describe my experience using it other than “rudimentary,” and a lot of that is on me for not really understanding the best way to set it up.
It's still frustrating as hell to come down in the morning, having given it a list of tasks to do overnight, with tests to pass before they're "done" and find that it worked for about 20 minutes after I went to bed, and decided that it would stop at "3am" (it wasn't) and "not do significant work this at this late hour". Like WTF ? You're an LLM. You don't sleep.
Bloody training data full of humans demanding sleep. I tells ya...
[0]https://pmarchive.com/guide_to_startups_part4.html: "In a great market—a market with lots of real potential customers—the market pulls product out of the startup... The product doesn’t need to be great; it just has to basically work."
Classic monopoly move by who?
Apple created MLX as an open source framework to allow users to run any open model locally.
MLX was just Apple's bridge to what already ran in other hardware.
I hope Apple can take all this cash and do some stability releases like they used to do, bugs around things like Family Sharing, the painful "update" to Settings App, etc could all use a lot of love.
It's true, most people don't run models, but being the default platform for running open weights seems like it has plenty of advantages right now. Just like sales benefited from developers defaulting to MacOS for most open source languages like Ruby, Go, Rust, and TypeScript.
64GB+ or dedicated 48GB (2x24 on GPUs) is IMHO absolute minimum.
- 16GB for the weights at Q4
- 9GB for the full 256K context at Q8
- 7GB spare for overhead and system.
The problem is that these Macs have 32GB of slow unified memory.
Edit: I'm thinking of a headless Mac mini, if you meant running it on the same machine you're using of course you'll need more memory, but LLMs are best served from a headless server so that's what I'd recommend.
https://www.canirun.ai (five months ago: https://news.ycombinator.com/item?id=47363754 377 comments)
32GB is not enough RAM. I don't even own a device with less than 36GB at this point, and that device I only have because my employer is being cheap. 64GB is a reasonable starting point for running local LLMs + normal tasks. 128GB let's you really run most smaller models like Qwen 27B and 35BA3B with good context. Even Qwen3.8-Flash-Next runs in 128GB with a 4-bit quant.
32GB would be limited to running models like Gemma4 12B and smaller dense Qwen versions like 9B unless you were using very small quants which damages quality of response.
Is this the best? No. That's why I said the sweet spot. Getting from 16GB macs to 32GB is perhaps possible. Jumping to 64GB or 128GB as the default is simply unreasonable right now.
A year ago you could get an M4 Mac mini for $399 on sale and now the same one used goes for over $700. The general AI RAM/SSD spike is part of that but there was also a huge demand spike for small, powerful, desktop machines that could be easily configured with these workflow tools.
1. Running an agent like OpenClaude. The $599 Mac Mini was an insanely good deal for this. I happened to buy a M5 Pro Mac Mini for $999 last year for other reasons. The equivalent is now almost $2000; and
2. Hardware for running inference on local models. This to me is the far more interesting market because Apple has a real opportunity to disrupt NVidia's stranglehold on the market.
With current architecture, the largest model you can reasonbly run is the amount of memory on the GPU and is a function of the quantization (eg int4, int8, fp8, fp16, etc) available and the number of parameters. NVidia aggressively segments the market. The most VRAM on a "consumer" card is 32GB on the 5090, which allows you to run ~31B parameter models.
In comparison, the RTX 6000 Pro has only slightly more CUDA units than a 5090 but has 80GB of VRAM. A few months ago they were $10-11k. Now they're ~$16k.
Macs use a shared memory architecture. Apple has previously sold Mac Studios with up to 512GB of RAM. Almost all of that memory can be used to hold much larger models without taking a penalty for interconnections between different GPUs or machines. Plus Apple interconnects between computers are actually relatively good by chaining TB5. It's still slow but it's about the best non-enterprise option available.
But the previous Mac Studios just didn't have the raw FLOPS and memory bandwidth. The M5 Ultras are up to 1.2TB/s of memory bandwidth. M3 Ultra had ~900GB/s. RTX 5090s and RTX 6000 Pros are 1.8TB/s. The current best HBM3 NVidia DC GPUs are at 3.2TB/s IIRC. But the M5 Ultra has a claimed ~4.5x the FLOPS of the M3 Ultra.
We don't have our hands on these yet but it probably means they are going to be much closer to a 5090. I expect ~50% of a 5090's inference speed. That may sound bad but it's actually really good because a 256/512GB Mac Studio can probably locally run the best Flash models. With NVidia hardware you'll need to spend many tens of thousands for that.
We'll see what the inference speed is but I expect it to be usable. DeepSeek v4 Flash, for example, will be entirely runnable. We're not at DeepSeek v4 Pro local yet.
I still have zero clue how "Buy a $599 Mac Mini to have a sandboxed LLM API caller" became the default. If you're not doing local inference and don't need to inject into iMessage or iCloud, all you need to run openclaw-style harnesses that call external APIs is a Raspberry Pi 4B, an N100, an HTPC, or that 10 year old laptop sitting in your desk.
And?
Local inference solves so many of the privacy and inconsistency problems with these frontier subscriptions.
One very efficient option today is to have the cheapest Jetson (Orin Nano) run the classical robotics stack, then have a base mac mini run nothing but the VLM. The Mac mini is considerably cheaper and faster at these workloads than the mid-range Jetsons.
I think this wonky situation is because Apple us under immense consumer pressure to absorb the ridiculous memory prices, while the Jetson is aimed at "business" and much more likely to fluctuate with the market. Last year I bought a Jetson Orin Nano 8GB for $375CAD, today that official nVidia Amazon page is out of stock and other sellers have it listed for $900-$1100CAD. Absolutely bonkers pricing.
I can't even name one person who fits this mold, let alone a non-insignificant proportion of people. Who are you thinking of?
[1] https://tech-insider.org/ca/steam-deck-price-increase-2026/
Is this Claude code? Or your local? I assume Claude? I'm more than a little staggered by this, like, it makes no sense! It doesn't even serve Anthropic's interests (surely better for them if it burns your token quota so you have to buy more the next morning.) The LLM just... decided? I'd be so mad.
WTF indeed. Can one even file bugs?
I think that's Anthropic trying to get you to not extract as much value out of that subsidized subscription as possible.
I'm not surprised at all.
Context: I have a farm of DGX Sparks and several RTX 6000's, and can run very close to foundational models with ~2 sparks
That said I have an RTX 5090, not a Mac Mini, so it's not exactly the same level of performance... The latest open models run at 200 tpm at around 30B params.
What you'll learn pretty quickly from said engineering is that there's a lot more to a good LLM than just the weights themselves. You need a good search provider (also self-hostable, but sounds easier than it really is). You need (well, it's debatable) a memory system. You need a good system for up-to-date library references like a Context7 (also self-hostable but the options are surprisingly not that good). You need a good set of specialized subagents that can perform various tasks well -- for the sake of "doing things well" but also managing context efficiently.
When you've got all that, local models can be _extremely_ useful. But there's one other important thing and that's decent hardware, unfortunately. A lot of people try out local models using small consumer GPUs or Macs and are rightfully unimpressed with the performance. And if the performance doesn't get them, usually they have expectations that they'll perform at Claude levels out of the box. Getting in that neighborhood, like I said, definitely requires some work.
I keep hoping that one day some comment is going to paste a link to some kind of idiot-proof guide or piece of software that’s “90% as good as Claude but running local.”
And by 90% I don’t mean that the model is 90% as good or runs 90% as fast, more like all the other stuff you mentioned is set up out of the box.
So yes, they are genuinely very useful, but they are not yet a full replacement unless you have more powerful hardware and or don't need more intelligent ai.
What's an RTX 9070? Do you mean the RX 9070 or RTX 5070?
What they are doing that you might not be is context caching, retaining a user profile about you in vector db, and adhering to strict output templates - often invoking multiple different prompts to assemble the end response to the user.
You’ll notice the mainstream ones always kinda answer in the same way. They are also able to mix code and images with text (multi-modality) because they are breaking down and routing your prompt to different models then assembling a result.
That’s all just software, has nothing to do with the model.
More people imo should be experimenting with it rather than chasing better models. This area is where all the differentiation is happening in the services, and what gives an AI provider like Gemini or Vidu or Singify an edge over running the equivalent locally
That is at best misleading and at worst outright misinformation.
Yes, the whole Deep Learning thing was luck, but as with most lucky things, they ensured they were positioned to capitalize on it.
This story turned out to be false but I think smart, reasonable people a couple years ago could have believed it with conviction. It doesn’t really seem like “completely asleep” to me.
But "oops, we missed that people are interested in AI work on our machines" seems like a really fucking big myopia. But then again, Tim's off to retire on a bed made of cash this week, so...
My vibes were that Apple wound down the “actual work” side of their operations (including machines like Xserve), because Ives couldn’t handle the unsexiness and unpredictability of business requirements in hardware.
He was self-indulgent and only wanted to work on things that “vibed” with him, rather than what the customers needed. It’s easy to be creative when you get to do what you want to do, it’s hard when you have hard constraints.
However , I don't think they expected the level of Enterprise interest they saw.
I've been using one for about a decade as a media server.
It just sits in the cabinet happily running the macOS TV program with the video files on an external hard drive. Playback on the TV is handled by the AppleTV's built-in Computer app. Works beautifully.
I have more movies and TV shows on that box than I could watch in my lifetime — a combination of ripped DVDs (Netflix, public library, and purchased) and OTA recordings.
When the cable goes out in my neighborhood (frequently), or a big storm screws up satellite reception (seasonally), I just don't care because I'm all localhost. As long as the lights stay on, everything is fine.
No ads. No privacy violation. No fees. No bandwidth congestion. No buffering. No subscription rate increases. All I pay for is electricity.
Frankly even the entry price is a bit high - I remember buying one for my son a few years ago (M1 mini) and it was a few hundred; now we're up to $900 for the base model.
The only reason for this huge speedbump is that chip makers have been dragging their feet for the last 10 years with "just enough" memory.
Stuff is crazy expensive.
The base price of a mini has only gone up $200 from $699 in 2020 to $899 today, and for $699 you only got 8GB of RAM instead of 16. Yeah the price has gone up but not nearly as much as you seem to be remembering...
What makes you think that? There's a lot of data centers that sell you access to colocated Mac Mini's, they have added FileVault unlock via SSH in the boot process which also makes things easier. There's not that many reasons to run a Mac in the cloud unless you have some very specific Mac related workload.
https://www.macrumors.com/2026/08/26/leaked-images-of-apple-...
https://www.reuters.com/business/apple-begins-shipping-ai-se...
https://wccftech.com/apples-private-cloud-compute-server-m5-...
https://forums.macrumors.com/threads/photos-of-apples-own-ne...
There’s articles about them, Apple uses them internally for AI services. https://forums.macrumors.com/threads/photos-of-apples-own-ne...
Plus they did sell rackmount servers for some time.
Here's a leaked / rumor image of Apple servers themselves.
https://www.macrumors.com/2026/08/26/leaked-images-of-apple-...
> [still claims there is no reason]
In my opinion, 98% of the work most devs would send to an AI can be capably achieved with a local model and a frontier-level model is overkill.
The goalpost moving feeds right into Anthropic and OpenAI's interests.
$200/month is vastly cheaper than owning and operating comparable hardware.
And all this before you get into privacy/security/compliance stuff.
I started tasking fable with huge projects over the weekend and now I hit at very least fable limit by monday.
What sorts of things are you doing with the local LLM? Anything interactive? Should I take another look?
And I don't remember to have been able to have pushed to 200k context Qwen 3.6. 3.8 is running on my RTX 5090.
EDIT to add that you need to reserve 8GB for the system if you don't want to cause problems on macOS, which means 32GB RAM = 24GB max for model + context. It takes 18-19GB to load a 4-bit quant of Qwen3.8-27B, so I'd be really surprised if you can actually get a 200k context window. You need to fit within a 24GB WSS (which is generally a more constrained RSS) to get stable performance on 32GB RAM.
Parent mentioned their $20/month subscription. It's definitely in Anthropic's interests for you to not use it.
Quantized Qwen 3.8 Flash Next could maybe run eventually on that card with a highly optimized inference engine that dynamically caches the hottest layer experts. Even then you run into some hard limits.
They have been taking NPU's seriously on their phones, tablets and laptops since the M1.
Then they enabled fully-connected RDMA for 4 x 512GB MacStudio's = 2TB RAM. Perfect for a large Mixture-of-Experts model.
It would be very strange if they didn't notice their product line had landed in a new sweet spot.
While the company I am in is embracing AI the disconnect and delay between what is available and possible versus what is approved and permitted is a three month window. The State employees I speak to are just now getting around to writing their usage policies for internal AI usage.
Prior to two weeks ago, I was just using Pi and Ollama.
I have tried my hand at putting together a few harnesses and I finally landed on what I like. Been working on this small app to handle running llama-server for me from any device that has the llama-cpp stack setup: https://github.com/SamInTheShell/loom
Qwen 3.8 is the first model I've been using that hasn't been having issues doing edit calls. Here are my llama server settings and GUFF that I use: https://gist.github.com/SamInTheShell/0bf838e8dc5093583b688e...
But in the meantime I get dishes done, vacuum, flip laundry... etc etc
Frontier models also seem in such a rush to emit anything they produce a mess that needs steering all day anyway
While I have not tested it, it feels like my local setup going slower is better at producing code that works the first time as its not trying to look fast for marketing sake
How is using an open source framework to run open models a monopoly move?
Works out really well.
Only if by "secret" you mean "announced in multiple press releases and a public event with federal, state, and local officials at its new sever factory in Houston."
https://www.apple.com/newsroom/2026/08/apple-opens-advanced-...
In _a_ cloud: their own virtual private cloud. They also have enough power to negotiate contracts with strong privacy provisions.
What privacy provisions would you want to add to AWS? Most of the reasonable strong privacy provisions you'd want are already there and/or available if you want to sign up for it, even including US Govt Top Secret data if you meet some approval.
If they violate their privacy agreement with me there's effectively no punishment I can get that they will actually feel. There would have to be a class action law suit and they always just settle those for some small amount and admit to no wrong doing
e.g.: previous crypto hype-cycle
https://www.pcgamer.com/nvidia-cmp-graphics-card-availabilit...
In general, I'm a big believer in doing more with fewer resources, within reason, and think having local setups really helps me be mindful with what's happening under the hood with these systems and managing context efficiently to get high quality results.
Yes, of course the most economical path is to hand over all your data and become fully dependent on a cloud provider who is already operating as scale, hoping that they won't change/remove models, hamstring capabilities, or raise prices.
If this were a thread about hosting your own email or blog or cloud photos, you'd have plenty of people out here telling you how easy it is to do it yourself instead of relying on Gmail for email or WordPress/Medium/Substack for blogging, or iCloud for cloud photos.
And yet, without fail, every single thread about self hosting local models seems to have some copy/paste form of this cost-savings argument.
Where is the appreciation for this cool thing GP built? Where is the appreciation for the desire to figure out how to host your own version of the incredible capabilities that were not available merely a few years ago? And why, on this site of all places, would someone advocate trading all of the knowledge and independence gained from learning how to host something like this ourselves in favor of throwing it all over the wall to Google?
Come on.
Because there are more privacy guarantees there, depending on the provider. "But what if they violate their contract!" is some pretty tin-foil hat stuff.
How is this any different than a business running their website out of the cloud, assuming you are using a provider with appropriate contractual terms?
You can care about tracking and ads but still be comfortable storing your backups in the cloud, and many have been for quite awhile, even sometimes without encryption - that is totally different than e.g. Meta actively trying to track you and understand your relationship graph and your purchases etc.
It will be horrible to be dependent on an AI who is also be trying to sell you various goods and services.
We're going to need AI whose loyalty is to us and only us.
There is no reason to believe that equivalent level model output will be more expensive in 12 months, let alone almost 4 years from now.
Of all the good reasons to use local AI (privacy, etc), worrying about not having access to cheap models in 4 years is not one of them.
It's almost never a drop-in replacement, and having to check and adjust integrations and workflows with new models gets old fast. My task was perfectly solved by the old model, I don't need a newer, "better" one - especially at higher prices ("more cost-effective" my foot). Local models lets one choose a model and freeze the downstream integrations forever, without being forced on the 6/8-month upgrade treadmill by aggressively short, scarcity-driven hosted model-deprecation schedules.
> They are things that I would not be comfortable sending a cloud provider
It's also an old machine that the commenter already has; it's intellectually dishonest to compare it to the price of a brand new, 4-iteration-newer machine.
a) model I pick will not 'suddenly' go away
b) I am sure my data stays where I want it
c) my inference mac can run other things if I need to
I pay for that.
But what if they violate their contract!" is some pretty tin-foil hat stuff.