My local model setup on an M4 Pro Mac Mini
lws.io
lws.io
I suppose I am waiting for AI-in-a-Box to come along so I can (painlessly) join in.
(I'm sure wrangling with all these esoteric aspects of LLMs though is fun for some people.)
Which are...?
1. Getting information (such as information about hardware unfamiliar to me) when not connected to the internet, which happens occasionally in my case.
2. Continuing to learn Rust by way of toy examples, puzzles, and comparing aspects of various solutions, for example from LeetCode.
3. Reformatting data, for example from a PDF to a markdown table, or converting receipt images to text.
4. Simple translation/explanation (e.g. I'm teaching my wife one of the languages I speak but sometimes may not know/have the words to explain the full nuance of a translated word).
5. Summarization. One of the webnovels I'm reading has some very boring parts I don't want to slog through, in those cases I simply make the LLM summarize that part and move on.
Etc., you get the idea. It's not unusable for coding, but it would make many mistakes when making a whole feature and the context lengths are limited to around 30k-40k tokens by my RAM. I could give it access to the web but I simply use an online model when I need that sort of thing, again partly due to the context limit.
Edit: The MLX version of Gemma 4 26b a4b does about 62 tok/s.
I would try a different version of the model from HuggingFace while ensuring it's MLX. I'm also using LM Studio, not oMLX, and I've seen some threads like these:
https://www.reddit.com/r/LocalLLaMA/comments/1spuwir/omlx_10...
What are you using them with/for?
That means all these data centers are being heavily utilized by actual end user inference demand. Well, some is research on new models, but a lot is actual end user demand. No one has given an explanation of why peoples usage would decline.
On top of that, margin on inference appears to be decent. It's model training that's a serious financial burden.
And maybe that's where there will be a slowdown, maybe the market doesn't justify spending as much on R&D as it does, but the end demand for inference is there.
Does that justify these stock prices? That's a different question. But the housing boom left behind endless rows of empty homes because demand disappeared. The 'dot com' boom left behind thousands of miles of dark fiber that'd been built out well ahead of demand for bandwidth. I can see the stock market having a giant sell off, but I don't see data centers sitting idle in that same fashion.
Here's what you have to believe:
- AI demand is at least several times larger than what can currently be satisfied, or will grow. (This one I can buy, but...)
- AI chips (GPUs, TPUs, compute-in-memory, whatever else is being studied) will not get significantly more efficient than they are now. It will not be possible in, say, 5-10 years, to do 2X or 4X or 10X more AI requests per rack than is possible now. I think this one's the single most likely thing to be false, since all computing history contradicts it.
- Edge devices (PCs, laptops, specialized but smaller scale AI compute nodes) will never be powerful enough to run frontier models at a reasonable price that's appealing for professionals, enthusiasts, or businesses, and there will never be a market for this. None of the demand will be served on-device or near-edge. AI must all go in giant data centers.
- AI models will not become significantly more efficient than they are now. There are no large gains on the table from better model architectures, better training, more efficient quantizations, better harnesses, etc.
If all those things are true, than the current planned like 4X-10X increase in data center capacity makes sense. If even one or two of them are not true, then the planned data center build-outs start looking excessive. If all four are not true, it's a total bubble that will crash and burn. Answer is probably somewhere between, but how far toward bubble? That's why I picked a number like "only 20% ever gets built." It might be as high as 50%. It ain't gonna be 100%. The planned built-out is batty.
Oh I forgot two more...
- Data center capacity currently serving non-AI work loads does not shrink through either reduced demand, more efficient software, or (most likely) faster chips and denser RAM. If that happens, more pre-existing DC space can serve AI work loads.
- Orbital solar powered compute nodes never happen. If this happens (free power! much less political opposition!) then terrestrial data centers have significant competition.
99% of the cost was in input tokens, I only used like 100k ish output tokens. It was a one shot task asking the agent to implement proxy injection to Guice. It did a pretty amazing job.
If you were to use hosted LLMs for a lot of agentic coding, a maxed out M5 Ultra Mac Studio would pay for itself in under a year.
Considering that I hit the 1M compaction multiple times per day with codex, it would definitely cost at least $5-8/day to use deepseek how I normally use codex.
He was a lead engineer, so after he announced it wasn't going to work, everyone pretended it never happened. But we all knew.
Also the entire purpose of them buying it was so the department had a LLM.
I proposed A6000. That ended up working.
With an 8x MI355x cluster at full tilt and including cooling, your power draw runs ~17kW. That's what it looks like when it's running full tilt. To be fair, hey that's pretty expensive. It does mean 8 multi-trillion parameter models unquantized running 24/7 without pause. And you get the full month like that, your monthly token limit is the time in a month. That cluster, the electrical upgrade, the cooling setup, and the electricity to run it all costs less in 2 months than your maximum affordance from Anthropic does in the same time period. Two billing cycles, and realistically it's more like two weeks. In 4 quarters you've wasted over a million. Like, what are we talking about here?
Now if you aren't using AI all that much, which is perfectly valid, and especially if you aren't using it at its absolute maximum, the story changes. Because even though at that point you're not paying nearly as much in electricity to run the cluster anymore, you still have the $300k+ capex to get the setup in the first place. But if we're not redlining it non-stop, then we're not really talking about performance anymore, are we? If your org never comes close to hitting token limits, it's probably because AI is rather marginal for you. Which again, is perfectly valid. I don't even use AI professionally.
Fact of the matter is, if your corp can justify the capex for a cluster and makes heavy use of AI, you are literally burning money by not having one in your building. The numbers are painfully obvious. Even deepseek isn't as cheap. This is before we get into things like LoRAs, custom inference pipelines, etc. which you know are kind of important if you actually care about model performance.
Pretty expensive is an understatement. You couldn’t buy one of these if you wanted to right now. If you could it would be multiple hundreds of thousands of dollars.
> It does mean 8 multi-trillion parameter models unquantized running 24/7 without pause
You can’t even run one unquantized multi-trillion parameter (>=2T) model on 8 x MI355x with enough context for concurrent users. I don’t know how you think it’s going to run 8 of them at the same time. Did you mean 8 concurrent sessions?
Your math is way off across this post. If replacing an Anthropic subscription for a whole company was as easy as buying a box for the office and then breaking even in 2 months, it wouldn’t be some little secret that we only discover in a comment online.
You can: https://www.exxactcorp.com/Exxact-TS4-149591758-E149591758 . You can get thousands of tps of GLM 5.3 output out of this thing, which grades around Opus 4.8. Payoff is around 1 year vs. spot prices on these GPUs, including power.
A system with 4x RTX 6000s costs about $60K these days, and can (as you note) trade blows with Opus 4.8 if not Fable. In fact, it'll give you a better pelican than Fable 5.1, and in less time.
Okay I love the open models, but the hype is getting ridiculous. The models you can run on 4 X RTX6000 are not Fable level.
And Opus is no slouch. I'm satisfied that GLM 5.3 is just as strong as Opus. Z.AI has promised/bragged that they will be at Fable 5.0 level by the end of the year or early next year, and I don't see any reason to doubt them.
No, you can get a quote for possibly being allocated one in the distant future.
The backlog for these is huge. You cannot buy one any time soon.
They're a good provider but you have to be a big shot buying NVL72s before you're getting anything within your payback period.
Also, the lead time I quoted was for individual 8x nodes.
Obviously, I quantified both the operating expense and the capital expense in my post. What I find curious is that you're quoting me talking about the operating expenditure, and changing the topic to be about the buy-in like these are interchangeable things. You don't think that this is a crucial and important distinction?
> You couldn’t buy one of these if you wanted to right now.
You could have spent all of 5 seconds of searching rather than just assuming[1]. You're not buying an Nvidia Superpod™.
> You can’t even run one unquantized multi-trillion parameter (>=2T) model on 8 x MI355x with enough context for concurrent users.
That's certainly fair a point. Although in the English language, especially in legal contexts, the multi- prefix is used inclusively for fractional values. That is it's strictly >1, not >=2. IE an 18 month contract is a multi-year contract, or a $1.6 million dollar asset is a "multi-million" dollar asset. But this is uninteresting semantics.
You are right, but it also doesn't matter. The gap is just that big. You can run 1 single user of Kimi K3 and still not even come remotely close to the $70k or so that a single Opus 4.8 user can burn over the course of a month on left on max. An honestly lowballed amount I know from anecdote. The per-token cost is just really expensive.
> Your math is way off across this post.
You made one technical point above, one that doesn't ever arrive at a relevant rebuttal to the substance of my post. But please, I'd love to hear you elaborate, especially because I didn't actually give much math at all.
If you want math though, here's the math. Let's say you are paying a ridiculous amount of money for electricity, a price nobody in the US pays -- $2 per kilowatt hour. That's about 5x the average rate in California, 4x as in Hawai'i. 17kW @ $2/kWh * ~8766 hours in a year puts that cluster's electrical costs at just shy of ~$298k annually assuming it takes no breaks. Let's make matters worse and round that up to $300k. It's also assuming you didn't invest in a solar hookup for your building, which I don't know why you haven't at this point, especially if you're installing a CDU for your new cluster. 12 months of Claude burning $70k a month is $840k. For a buy in of, you know what, let's call it $500k. Why not? It still doesn't matter. The operating cost is so much lower it's paid for itself plus an additional $40k in the first year. Even at a ridiculous penalty in electricity that nobody pays, even overinflating the amount of money you'd pay for the cluster and the infrastructure to get it set up, it's not even remotely close for a single user where the gap is smaller (IE, you're not wasting "a million dollars" in a year by maxing out the $200k scaling limit every month)
You can of course trot out the point that oh, in 12 months this setup will be extremely outdated! It doesn't matter. If the work it was doing today was useful, it will be useful next year too. And with the rapidly encroaching diminishing returns from parameter scaling, you're probably going to be just fine for a while. Maybe grab a quantized version of a newer Chinese model at the end, before grabbing a newer generation of AMD node. Those MI400s are looking pretty sweet after all.
> If replacing an Anthropic subscription for a whole company was as easy as buying a box for the office and then breaking even in 2 months
If you're locked in, then you're locked in. But don't pretend like you're saving money. You're not.
> it wouldn’t be some little secret that we only discover in a comment online.
Why does this have you so nasty and defensive? It's not a "little secret" that running your own infrastructure is cheaper. Of course it is. You know what else is cheaper? Owning your own office building out in the sticks, rather than leasing part of one in the city. Not everybody can make that work, there are no free lunches after all.
History repeats, these same exact lines were rolled out ad nauseum during the cloud craze. Datacenters are businesses, not charities. Frontier companies rent quite a fair amount of their infrastructure. Even if they resold that compute below cost (they don't), there's a pretty steep cliff before the economics start to look attractive.
[1] - https://www.avadirect.com/GIGABYTE-G893-ZX1-AAX4-Dual-AMD-EP...
I guarantee this will not ship to you any time soon.
The current lead time on these GPUs in measured in years. If you didn't place an order for this a long time ago, it's not coming this year.
Being able to add it to an online configurator does not mean anything right now.
> 12 months of Claude burning $70k a month is $840k
Your math is completely useless with these arbitrary numbers pulled out of the air.
If you want to begin calculating payback period you'd need to look at token costs, cost per task, utilization rates, and so on.
> The operating cost is so much lower it's paid for itself plus an additional $40k in the first year.
You went from paying back in a couple months to paying back in a year but you still haven't even talked about tokens or concurrency.
You're also neglecting the fact that hosted tokens are going down in price at a rapid rate. If someone was paying $70K per month in tokens for Opus this month, that same level of compute is going to be much cheaper 12 months from now.
> Why does this have you so nasty and defensive?
Not nasty or defensive, just tired of these armchair claims that it's easy to go out and buy an 8 X MI355X box from people who obviously have no idea what the hardware lead time is like right now, or who haven't considered the actual math on token costs and payback times. You're still making a lot of claims without a single discussion of cost per task or token.
And I'm not even considering their time spent fiddling, fine tuning configs to adjust for ram, updating/benchmarking models, etc. Which is probably more expensive than the mac so the math is even more wrong.
The assumption, the starting point, is that you have a line on the hardware. Asking around, some distributors have a 6 month lead time on Instinct GPUs, which curiously enough is about how long you'll be twiddling your thumbs waiting for the cooling loop to be put in. Yes things take time.
> Your math is completely useless with these arbitrary numbers pulled out of the air.
Your dismissal is worthless if you can't even be bothered to provide a counter-example. You've not provided a single iota of quantified reasoning beyond my original not accounting for the space used for the context of concurrent users.
> If you want to begin calculating payback period you'd need to look at token costs, cost per task, utilization rates, and so on.
Now go back and carefully reread my original post. Yes, if you are not actually redlining an LLM for a billing cycle, the capex starts to be way more relevant for this setup. Otherwise, our constraint is time and our unit of measure is $/hr.
If you want to compare token cost, it may shock you to learn that Kimi K3 without speculative decode on this setup is slightly under twice as fast as Opus 4.8 max. That's still true when fast is compared with K3 with speculative decode, and now Claude is twice as expensive as a base rate. Oops. We're already burning more money over a period of time, looking at tokens we're screaming even further ahead.
> You're also neglecting the fact that hosted tokens are going down in price at a rapid rate.
Cool. Call me when Opus 4.8 max is $0.50/million. In 4 years you could have bought the 200 acres of land down the road from your building, started a 5MW solar farm subsidiary that you'll expand over time, and as soon as your connect is up, dropped the opex of the cluster down to its maintenance costs. That subsidiary will pay the loan required to spin it up back irrespective of your primary business. When you own your own shit, you can play your own game, stack your cards deep. Have a little bit of business acumen. Fuck what The Valley is doing, that is an ecosystem fully enslaved by economic nihilism, money isn't grounded there.
> Not nasty or defensive, just tired of these armchair claims that it's easy to go out and buy an 8 X MI355X box from people who obviously have no idea what the hardware lead time is like right now
This motte-bailey routine is both nasty and defensive, particularly when you keep prosecuting a geist of numeric justification that never arrives. All I've gotten from you is vague dismissals, one borderline irrelevant technical argument, moving goalposts and missing the point. Granted, not as egregiously as other people in this chain thinking we're talking about running 100B models on a Mac, I'll give you credit for that. But this whole time, we're just talking past each other. You make realistic points and I try to bring you back to context, but you have to work with me here too.
The point was that these companies are not selling to you below cost, they're not even selling to you at-cost. Just use your head. Venture capital isn't a magic wand. Frontier companies are in the red because they're in non-stop expansion operations at massive scales. Anthropic has an operating profit of half a billion dollars[1]. They are not selling you API usage below cost.
[1] - https://www.forbes.com/sites/jonmarkman/2026/08/17/anthropic...
But interestingly still extremely valuable on the second hand market.
The capital expense isn't the amount laid out. It's the rental cost of obtaining that capital, less the depreciation on the fixed asset over the period in use.
Going back to the OP, Apple gear is well know for having good resale values, which means the capital outlay isn't anywhere near as much as some people think.
If you just ask "who spent more in the first year" (100% depreciation) then even with 5-6 max accounts, buying HW will be a couple of times more expensive. But when does it make sense to ask that question?
Maybe the SotA models will need better hardware so your investment will not be useful after a year or you'd need very expensive upgrades? But then (as in Fable case) subscribers need to spend more too.
Idk where you live, but where I am running the M5 Ultra Mac Studio at max rated power 24/7 for a month costs C$42.
The considerations against Apple hardware are 1) hardware advancements 2) early access to the best models. But it’s really not that clear.
(The other guy who thought hosted models on openrouter are cheap has spent $100k in 5 years.)
Hardware is not magically getting more memory or bandwidth.
Believing there will be some magical optimizations to compensate for it is just dellusion.
Also, from what I can tell, MLX inference is not as well optimized as CUDA, and the M5 Ultra has additional kinds of AI compute which is unavailable on other M models. With the massive 1.2 TB/s 512GB Mac studios coming out, I think MLX will get a lot more attention.
In short: Todays models should run faster next year, and next year's models should also be more efficient.
You don't need frontier models to summarise or create an email.
It isn't, the cost is included in your electricity bill, not even talking about the cost of your time to set it up. It's very possible that it costs you more than a cloud mode would, you just don't want to calculate it properly.
...which is almost always true in a single request/reply mode and never true in batch mode. Single request usually 2x-3x more expensive than cloud and batch mode 2x-3x cheaper. Now, for narrow tasks, a finetuned tiny 8b model would dramatically outperform SOTA frontiers for a fraction of price, esp. on energy efficient hardware like Apple.
If you had only batch inference and enough of it to fill the compute to 80% then you get cheaper local models.
Local models can absolutely run in batch, what are even talking about?
> If you had only batch inference and enough of it to fill the compute to 80% then you get cheaper local models.
Even if you ran sequentally, single session, a _finetuned_ tiny (8B) local model on narrow tasks would abolutely mog SOTAs, any of it - Fable, Opus, Sol you name it.
I think the point was that if you aren't running your local machine at 100% for 24 hours a day then a cloud - with multiple clients - that is, will be more efficient.
a machine like this is about a years rent for most people.
a small car for most others.
48G RAM is pretty useful if you want to run k8s locally for tests / exploration
you'll also notice these articles rarely specify their context window in tokens, because it is small, usually 30k to 70k tokens and it gets slower as it fills up.
Was because I am back to using Linux as my workstation.
My Mac Mini is now a headless server for llama.cpp.
So, you are right that for these workloads , I would not be using the Mac Mini for k8s AND llama.
Another thing going against using a Mac for Linux containers is that there are no solutions that I know that properly manages memory : memory is given to the Linux vm , but never fluctuates if the needs in the vm are less than the initial request.
I know Orb Stack does that but is it proprietary. I think UTM does it , but not sure I would use UTM instead of Lima, Colima , multipass , etc to run containers.
That can be done on hardware that quite a lot of people basically just have and don't use 24/7 to the max - because it is their gaming machine or their programming and compiling workhorse, for example. Of course you are paying for additional electricity but even with napkin-math instead of a "proper" calculation, you are unlikely to pay more for running your own instead of something commercial (and that can be offset further with some of the "modern" electricity contracts and/or PV and battery storage). Especially if we are talking about a stack that runs most of/all the time when you are not using your machine and makes LLM calls regularly while running.
The work in software/admin to get whatever you want set up is similiar no matter which infrastructure you use.
Acting like an extra $20 on my electric bill is equivalent to a $200/mo subscription is... a take.
That said, it is really cool to be able to run an LLM on eg a Mac laptop. Just not a better experience on almost any metric for interactive use than eg Claude Code, beside privacy and guardrails.
How's the actual performance of Qwen 3.8 27B? On deepswe it supposedly performs slightly worse than gpt 5.6 luna high[1], but I can't help but think they've been benchmaxxed.
[1] https://deepswe.datacurve.ai/, https://unsloth.ai/docs/models/qwen3.8#benchmarks
A friend and I were actually discussing today how benches show Luna Max at about par on coding with Sol Medium, but how it's nowhere near in reality. We were speculating that maybe it's because a lot of benches are best-of-n, and should probably be worst-of-n, because variance in performance is killer with large coding projects. Consistency is what lets you actually build on this stuff.
Thank you. I just changed my opinion on this thanks to you. I agree now, since we tend to execute LLM tasks once instead of N times anyway.
I suspect models like Fable executes the same task N times in parallel and picks best answer or merges them to for a better answer.
Great showing from Sol, though.
But also, it's Baba Is You :-D
Which already comes from Claude itself. Clearly, they don't want to train on their own product.
If you are work from home and do dishes between prompts you can get a gpt3-like result.
I found it useful when I was... Well I didn't find it useful. But an Nvidia 3060 let me ask unethical questions pretty fast.
GLM-5.3-Flash at true 8-bit: 341 GB on disk, 328 GB resident, 288 experts across 46 layers, loads in 65 seconds. • 18.7 tokens/s generation, 35 tokens/s prompt, on a desk, on a $0 per-token bill. • Runs beside our whole agent city on one box with ~130 GB to spare. • Review test: caught 6 of 6 planted P1 defects, zero false positives, same score as the frontier model we pay for. • CRM test: 11 of 11 required records extracted, zero wrong writes, 45 minutes, first local model to clear the bar. • Serving a 131k-token window today; the model itself supports 1,048,576. Widened to 4 concurrent slots and still have 50gb+ of excess ram.
granted my cto still isn't moving all of our inference to glm5.3 but we've identified 40%+ that is currently handled by fable that we're routing locally instead and will do concurrent requests to verify/compare responses for a while.
$0 per-token bill
Hosted GLM 5.3 flash is like $0.15/mtok in $0.50/mtok out
I see Apple is currently selling a 256GB M5 for about $10K, so buying October's 512GB one could be, what, $13-14K?
A $0 per-token bill is great but this is clearly not something for normal people, just some businesses.
On a Laptop with 32 GB RAM and Iris Xe integrated graphic card, I get between 11-18 Tokens/Second with Qwen 3.8 27B and llama.cpp with sysl Intel optimisations. Same results with the vulkan back end, although sometimes it ends in weird segmentation faults due to the memory consumption.
With all due respect, I'm not clear why you are so surprised ?
By your own admission its a little mini-PC with 16GB RAM, I'm not sure what miracles you were expecting ?
Its a bit like complaining Rasperry Pi performance is terrible when trying to compile the Linux kernel.
For more complex or important tasks, costs, autonomy and privacy matter, but then so does performance/quality.
So I'm not completely convinced it's really worth it; but it's tempting!
The models used in TFA are halfway in between the traditional "free as in beer" software. Open weight means once you download it, it continues to work forever; and you can also do your own RL on them; but you can't really see what went into their training, nor train a new one yourself from scratch.
And that's probably good, otherwise the free ai would just be unavailable for everyone else
I’d rather use a tool where I know the limitations, over a tool where the limitations and strengths keep changing.
This way I know where in the process I ought to step in and pay attention.
There are 2 main reasons for running local LLMS.
1. Process private data/work with uncensored models.
2. Use a large amount of inference that would quickly blow through rate limits and/or run up API costs.
The thing that is critical for 2 is that a) you have to have a sweetspot between a pretty good model, which means largest parameter counts, and fast enough token generation where you can run agentic loops. The latter is needed because you aren't going go get the "intelligence" of larger models to form shell commands and run tools to figure stuff out, so the only way around that is to have custom agentic loops to force the model into doing what you want, which results in more text processing.
From my testing, Gemma4:31b is basically the only local model that can be relied upon to produce accurate results. Qwen models chase benchmarks, which results in MoE models (thus the A3B in the model, i.e 3 billion parameters are only active during inference). In general, these are good for very specific tasks, but fail to be accurate in considering cross task data, whereas Gemma, being fully active does a much better job. If you only need to do a very specific deterministic task, those models are pretty good.
As an aside though, if your task involves pure text processing (for example take html data, make it into a markdown document), you can also additive train Gemma270M quite easily all on CPU, and on a decent CPU it gets like 50-100 tok/sec, no need for any extra hardware.
The thing with Macs is that while they can run those models and larger models no problem, the tok/sec is very slow. This limits effectively what you can do with the models. On the M4 that the poster mentioned, Gemma:31b will run about 20 tok/sec. That means that when you wants to write a whole code file or process large context, you have to wait for it to do things. Compared to workflow with larger models, where file generation often takes like <10 seconds, it takes a while to adapt.
The only benefit of using Macs is the price for Mini and cheaper studios. However, once you reach the total cost of about 2.5k (note that the M4 statedin the article us about 2k), building a gfx card rig is the way to go. You can get 100 tok/sec on a 3090, and it will feel a lot like the cloud models.
I have a beefy Linux box with a 4090 but never took the time to set it up properly beyond simple testing; any tutorial you would recommend?
If you want to build agentic frameworks, use llama.cpp with its built in http server, and build the framework with python
---
Qwen3.8-27B-4bit, Prompt Processing (PP) 66.3 tok/s, Token Generation (TG) 11.8 tok/s
Ornith-1.5-35B-A3B-MLX-4bit, PP 379.7, TG 45.8
Ornith-1.5-35B-A3B-MLX-4bit, PP 381.5, TG 46.4
Qwen3.6-35B-A3B-mxfp4, PP 389.6, TG 47.6
Qwen3.6-35B-A3B-OptiQ-4bit, PP 342.6, TG 44.4
---
Qwen3.8-27B-4bit generally runs out of output token before completing the task though excellent partial results.
Ornith-1.5-35B-A3B-MLX-4bit seems to get in the loop often specially with tool calls.
Qwen3.6-35B-A3B-mxfp4 seems to be optimal with speed and quality output.
I am going to test Qwen3.6-35B-A3B-4bit soon with same code block just to check my intuition that any derivatives don't seem to perform better than the originals.
Finally, I settled on Qwen3.6-35B-A3B-4bit with 32,768 context window and 16,384 max tokens.
---
Additional results from Qwen3.6-35B-A3B-4bit (Can't edit previous comment)
Qwen3.6-35B-A3B-4bit, 329.7 PP, 41.3 TG
I got 400 pp tps on a 10k token input. Your numbers seem suspiciously low, maybe the input was too short to measure properly? And this dense 27B is slow, the MoE A3B models get to 1000 tps.
I've since acquired two DGX Sparks, and it feels so much snappier.
the sparks have much slower memory bandwidth is the trade off
Another benefit of the 2x spark setup is that you can parallelize to ~6 streams pretty efficiently.
All depends on the workflows you’re using it for.
I’m quite excited for the M7 class machines.
Yes but it's easy to replace them.
The main reason should be privacy.
Meanwhile the stock market has Nvidia at the top... Until everyone gets cuda.
There isn't really an overlap yet.
Individual Nvidia cards exist on desktops but they're not really oriented for regular inference so individual developers are left with Macs or datacenter resources as their options.
I, personally, do use frontier models in the cloud for a lot of (meta-)cognitive analyses that are heavy enough to have me run against the limits of payed accounts regularly - so I'm neither a Luddite nor stingy with cash in this case.
However: I have pretty good experiences with local models as well. My solid but hardly extreme desktop (with one RX 9070 XT 16GB) mostly serves gemma4:12b and specialized models (embedding) to my local network. This is for general use like simple queries, simple code, reformatting and the like but also for two specific tasks that are permanently running:
a) It's connected to Home Assistant (as a second stage after very simple "turn light XY on" commands which get processed without LLM). So, I can mumble into my smartwatch "computer, how much gas do we have in the warp core and how much energy did the bussard collectors make from the cosmic dust today?" (or describe a more complex light scene or create an automation I want or whatever). The phone transcribes that - with a local model on device - and fires it to the desktop who has agentic access to HA, looks through the sensors and data, sees that I've tagged my solar panels and battery with nerd vocabulary. It makes the right conclusion, converts a few units and gives me back a nice overview. All hands-free while I'm sitting on the toilet.
b) It's the LLM backend for a personal radio station run by a fleet of nerdy/quirky AI DJs who's archetypes are represented more than well enough in the latent space of the "small" model to produce funny results. The DJs can produce consistent, individual segments and programs, run a playlist that works well for me (based on multi-layered audio analysis that also uses local LLMs), respond to song wishes and generally produce much better recommendations than Spotify ever could for me. And you can also put multiple of them in the "studio" to create hilarious crossovers that you would not get from a commercial entity because the IP owners would rather shoot each other in the face.
All of this doesn't even max the available resources, so I can shovel F5-TTS into the VRAM as well and have all my DJs have good, locally created voices (or voice clones of Captain Picard and Han Solo, if I wanted to) based on zero-shot voice cloning.
--> Far from "unusable". It just depends on the task. And I neither have to hand my keys to the Navidrome server nor to my Smart Home to any entity outside my local network.
My understanding of using a mac mini for ai (ie running a claw bot or whatever) is to have it 'always on' and a better price/performance profile than a cheap vps.
Is there performance (silicon processor?) so unique? As I guess it's not their graphics units. I see tonnes of people using mac minis for AI, to the point it almost became a meme.
Edit: yes I know this article is about local models, my question is a bit more general.
Apart from that, if it doesn't work out you still have a Mac Mini, which in itself is more desirable for many than a DGX Spark or Strix Halo if you have no AI use case.
Not completely true. It's memory AND memory bandwidth. You can have 1tb of memory but if you have awful memory-bandwidth you'll also have slow tok/s. A3B helps with this, but so does MTP.
From my experience, you'd be better off running the dense 27b-mlx with MTP than the 3.6 version with A3B. You say your model is ~20GB of ram, but the 3.8:27b-mlx is 18GB and gets me very reasonable tok/s, and greater speed if you disable thinking when not required.
In this case the 35b a3b makes sense as it has a PP speed of around 800tok/s
These are really good models but the harness has to be built around them. I have a ton of generated system prompts for specific purposes. Even parts of a SolidJS stack, for example Route management, has its own prompt. These are experiments but the results are real. If we build harnesses around small models, we can build a locally running WYSIWYG editor which works on plain text prompts.
The performance, in simple tokens/second, is not the most important factor. For many private data points, like emails, I would rather have a local graph based search and LLM on top where the harness is specific to problems like calendar, contacts, finance, etc.
I run all experiments on an 16GB M4 Mac Mini but coding agents building the harness are a mix of Codex, Claude Code and opencode.
Literally today, but it feels like an improvement.
https://x.com/mkagenius/status/2093730391429685732
(xcancel seems to have received a cease and desist)
I wanna get a desktop Mac for local ai so that I don’t turn my laptop into a delta 15k rpm fan when I run things.
I guess I’ll get in line for one hah.
Is that supposed to be hallucination? The human or other kind. Feels like a made up URL. It's .ai, isn't it?
If there was a "Mullvad of GPU clouds", would that solve the privacy concerns?
The problem I have with these is that the guarantees aren't strong enough (Phala, Near) or the models are old (Tinfoil). Chutes is mostly pretty good (cryptographic security all the way to the GPU) but I'm not sure it's possible to cryptographically verify the precise source code they run on the mode.
Phala isn't verifying all the way down but NEAR is and I know the CEO
If you look at [0] (the code they run in the CVM), there are a couple of things that worry me:
- They ship logs out of the CVM and worse, they send them to third parties (DataDog). Even if we could verify every bit of code running in the TEE, it's not enough to know the code doesn't maliciously ship prompts to a third party, we also need to audit what each binary logs.
- SGLang, the core inference engine, isn't reproducibly built. We have no way to verify that the thing they call "SGLang" is what they claim it is.
Really, it's the log shipping processes that worry me the most. Ideally, NEAR would minimise how much auditing needs to be done by having the minimal open-source proxy be the only thing with network access, making it much easier to audit potential exfiltration routes.
[0]: https://github.com/nearai/cvm-compose-files/blob/main/prod/G...
The way you are using it uses the internet and datacenters. It is costly to the environment.
Running locally is a significant cost savings in comparison.
Not many people share setup with actual setup handholding so that was very G of you
Agents require at least DeepSeek pro and even that is the minimum.
You might be able to get a good model to write instructions and run it in smaller models.
Otherwise, cool your AI got the current weather.
The Granite 4.2 models which are just recently out, are optimized to handle agentic workflows.
For local models, it's about using the right model for the right job.
There will always be a reason to run frontier models, but local models are well at levels that assist with stuff that don't need that level of complexity.
But I guess a $2000 Mac is probably better if you don't care about cost or quality.
Or I can use the Mac I already have.
Your example though, Ouch!
~8B Q4. That's around 5-10 tokens a second. Base M1 16GB mac would do 15-20 tokens a seconds. That's a 6 year old machine.
You do get what you pay for it seems.
Sorry local models are basically useless outside chat, I didn't even consider it.
You could run one of the smaller Gemma models to have a chatty Wikipedia.
This is the burning question for me, what are they doing with our hard work.
I'd have thought that sherlocking a user's $10M business would be too high risk, given the billions at stake if real evidence of this happening was found.
However, OpenAI are currently being sued by Apple for trade secret theft, and the way it was done seems to be abundantly idiotic.
So I'm torn.
I am curious is what is the 80% request served by this setup, I was using it for OpenClaw which run serveral cron jobs that discover stuffs over the wide internet, check my support system's unanswered tickets, browser X and some social media for me to filter the valued ones(though I have to say even with GPT 5.6 sol, the quality is low for the timeline X sent to me)
Btw, Tailscale is quite cool and did a good job, I was using it to serve the local LLM and connct the openclaw on a Linux Machine to it.
I read those blog posts to remind of wealth gap i have with the average hackernews user.
Thank you for the encouragement, i will work harder to reach your level.