No mention of the performance of the models? I'm able to load a bunch of different models on my little mini-PC with 16GB RAM, but the performance is terrible. I always wonder what performance people are getting with local models that they find is acceptable?
Some examples (keep in mind this is all indefinitely free for me, no burning quota away):
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.
My perf sucks compared to yours. Added it to the post - same model averages 325 tok/s in processing prompts, and 34 tok/s in token generation. What am I doing wrong..?
It’s not. Do it as a hobby or for privacy but for performance just use a frontier model api. You’re paying less than cost for something that would take tens of thousands to set up locally.
It does make me wonder how the hosted stuff is so cheap. For pretty much everything else, hosted/rented is more expensive but offers better convenience and flexibility. But for AI, even if you consider the total lifetime cost and are utilizing it heavily. You never break even by buying.
Qwen is weirdly expensive. Deepseek v4 flash is dirt cheap. You'd need at least 128gb of ram to run this model and in my experience, a days work with it costs around 80 cents.
So I ran the math, assuming the agent takes 75 turns per 200k context, with deepseek v4 flash it costs around $2.57 to reach 1M context in 375 turns. Cached input costs scale quadratically with # of agent turns.
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.
I've been hosting Qwen3.8-27B myself. On my endpoint it's $0.30/1M in, $0.10 cache, $2.03 out - so those agent turns that re-send the same prefix get a lot cheaper when cache hits. UI at inference.tiyuvta.ai/app if you want to try it. Hosted is up to 210 tok/s and 280ms TTFT with reasoning off.
That's not even remotely close to being true, even once you account for capex. You have to look at the actual usage, look at the token limits. Even if you're paying Anthropic $200k/month for scale-tier, you're going to blow through your token limits trying to run max output 24/7. Three users running Opus 4.8 at max non-stop will probably clean your monthly allowance from daddy Dario in less than a week.
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's not Claude at max output running 24/7 a month expensive though. More like, double your energy bill expensive. For me, that's ~$180/month without solar, and again that's running full tilt. That means 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. 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 ~$180/month 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.
> 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.
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 couldn’t buy one of these if you wanted to right now.
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.
Ah thanks for the solid info, too bad. I'd seen them come up as a pretty good price for 6000 RTX's in the past, which seem generally pretty available, good source for those?
Yeah, they're good source. But the price for those GPUs is 5 figs even with the nvidia startup program nowadays. Also, I went back and looked. Most of my GPUs are actually from Central Computers who were great, but Exxact is real too. So "lots of" was inaccurate.
Also, the lead time I quoted was for individual 8x nodes.
> Pretty expensive is an understatement. [...] If you could it would be multiple hundreds of thousands of dollars.
Obviously, I quantified both the operating expense and the capital expense in my post. What I find curious beyond that is 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. The per-token cost is just really expensive from an API.
> 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.
> 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. History repeats, the same crap was 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.
> You could have spent all of 5 seconds of searching rather than just assuming[1].
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.
It usually boils down to people trying to convince themselves that keeping their macs hot and with very little ram to spare only to get sub 50 tokens per second on a subpar lobotomized (quantized) model is worth it.
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.
> I guarantee this will not ship to you any time soon.
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. We're already burning more money over a period of time 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.
> 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.
"You can of course trot out the point that oh, in 12 months this setup will be extremely outdated! "
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.
Here’s an experiment: purchase an anthropic pro max subscription for $200/m. Now go buy the hardware to run DeepSeek’s equivalent. In a year, who spent more?
In normal times in which hardware used to depreciate (lately that's not the case and HW even appreciates, but let's not get distracted), if you calculate only with depreciation costs, plus the fact that when you have such a setup, it'd take many 200$ subs to cover your lack of limits in the other, I think it'd not be a clear victory for any side.
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.
It’s not so clear after 5 years that you’ll come out ahead. You’ll have spent $20k. The apple computer owner will probably be running local models that are better than today’s frontier on the same hardware.
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.)
Open weight models have been getting better/smaller every year.
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.
I run a similar setup to the one he described on similar hardware. I run bifrost and llama swap though (tailscale rocks). My local model usage is for some out of band batch processing one of my personal apps uses. Basically a personalized recommender for media, it curates stuff for me based on a database i've compiled over years, so non-interactive. For that use case, I don't really care that it might take a few minutes to run. It's free. The machine is just sitting there anyway. I have tried using qwen-coder and opencode on my M5 Max 128gb and compared to claude code it's painful. I did setup a workflow where claude plans, qwen executes (unattended overnight, again b/c it's slow) and then claude reviews. I benchmarked this several times and I ended up using MORE tokens with claude because it had to 'fix' all the qwen issues. While the code it produced was 'good enough' the fixes were worth it so I just stick to coding task using API models (codex and claude).
Can you share a bit more about your bifrost and llama swap setup? I’m facing memory constraints and am looking for a managed model solution that will help with hot swapping loaded models and stay-warm concurrency. Ideally with prioritization.
What do you want to know? Just start llama-swap with the models i have downloaded, add llama-swap as a provider in bifrost, expose the models you want and they become available in one single endpoint you can use in anything like opencode, openwebui or anything that speaks openai.
Yes, that makes sense. Some of my models currently run in ollama while others require their own inference servers. I’m curious about custom inference servers in bifrost and the ability to orchestrate keeping some models warm in memory while evicting and cycling other models. All of which span different providers in bifrost (I think). Obviously I can get the mechanics from an agent, more wondering about any experience with something similar.
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.
> It's very possible that it costs you more than a cloud mode would
...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.
true but if you're actually running k8s and similar workloads, chances are it might eat memory that LLM requires.
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.
I actually have a Mac Mini M4 Pro with 48G. I gave the k8s example because this is what I was doing with it.
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.
It sounds like they are doing something similar to what I described in my other post below. Personal media station.
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 regularly while running.
The work in software/admin to get whatever you want set up is similiar no matter which infrastructure you use.
>I honestly wouldn’t bother with local models right now unless I either had a 5090 and was happy with running Qwen 3.8 27B
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.
i have a 512gb ram m3 ultra mac studio setup with a gas city that runs one of my companies. today was the first time ever that a local model (GLM5.3 8-bit) was able to match fable5 in our tests.
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.
You still have electricity and capital investment. Envelope math suggests cheap electricity is costing you something like $0.50/mtok and the opportunity cost on the capital tied up and lost in the unit purchase and resale is going to cost you something like $2/mtok at 100% utilization (so, frontier model prices or higher at real utilization), and you don't benefit from any elasticity.
Hosted GLM 5.3 flash is like $0.15/mtok in $0.50/mtok out
Time to completion also must be considered. If I have to wait around for hours for a prompt to complete locally and I’ll need to iterate quickly, I’m better off hosted than local. If it’s “free” and slow it may just not be worth it.
Props to your parent commenter for including context size. Because 131k context window is prohibitively small for my coding workloads so I know a 512GB Mac won't cut it.
Are you using the right configuration for your own CPU?
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.
I took this thread and summarized it with Qwen3.6-35B-A3B, it had 1400 tps prefix and 60 tps completion. Very good performance. Using oMLX on MacBook M5 Pro 64GB.
I'm the author - hello! Added to the post! Qwen averages 325 tok/s in processing prompts, and 34 tok/s in token generation. That isn't instant, but it's quick enough that I never really think about it.
yes, share performance, numbers if you can, also i wonder if you figured out a way to do a 2way audio with local models, or even explored that. I have a very similar setup but not too happy with the token speed, will try omlx though !!!
Have a macmini m4 32G, not the pro version, previously everytime I tried local LLM is a bit disappointing, and I finally decide to not waste time and perhaps in the future invest a better hardware to server more modern and dense model
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.
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.
Apple is working from the 'desktop' up to beefy servers with 64GB+ RAM. Nvidia is working from the 'datacenter' down to beefy racks with terabytes of RAM.
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 experiment a lot with local LLMs, particularly small ones like Qwen3.5 4B and 9B. I have build multiple experiments to make harnesses that use these models for code generation, planning, local search, etc.
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.
Most people running local models would probably love to run larger models if only they had access to big enough hardware. I'm curious: to those of you running models locally, if there was a way to inference the model of your choice at a reasonable cost by effectively time-sharing a B300 rack through some privacy-protecting intermediary, would you consider that?
If there was a "Mullvad of GPU clouds", would that solve the privacy concerns?
runpod.io is essentially this. You can rent the hardware for cheap in small time slices. I do this whenever I need to do a lot of embeddings, fast. I have an agent skill that will estimate the optimum hardware to reserve for the time/price constraints of the job, and you can spin up temporary inference for cheap via their API as well.
yes, and it's already some offerings like that but they all cost a lot because they only good for "I have some idea of workload for N hours or days" lets rent it and run. That fine for some experimentation but if you think about renting something 24/7 even for example to share it with the friends that will cost at least 4x from any API prices as result (something like rtx 6000 48gb will cost ~$470/m).
Chutes, Near AI, Phala and Tinfoil all offer various privacy assurances around inference. Some of the bigger providers also offer "zero data retention".
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.
NEAR AI does a lot of things right but with this kind of thing, it only takes one mistake to completely break security.
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.
I really like these show and tell style posts. I’m always curious how people have their setups and what tools they use. Also the blog has a nice theme and is easy to read.
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.
Recent performance data on my M1 Max 32GB MacBook using oMLX. I have been working on identifying suitable model and config for my use case and system. Using a refactor and suggest improvements prompt for a specific Django code block using VSCode Cline extension.
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.
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.
Are they selling our data is the burning question for me.
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.
> The main reason to run local: cloud APIs are rented land. They can change their pricing, hit your usage limits, or swap the model being served behind the scenes whenever they feel like it.
190 comments
[ 0.64 ms ] story [ 9.1 ms ] threadWhich 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.
What are you using them with/for?
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's not Claude at max output running 24/7 a month expensive though. More like, double your energy bill expensive. For me, that's ~$180/month without solar, and again that's running full tilt. That means 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. 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 ~$180/month 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.
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 beyond that is 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. The per-token cost is just really expensive from an API.
> 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.
> 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. History repeats, the same crap was 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. We're already burning more money over a period of time 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.
> 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.
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.
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 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.
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
https://quesma.com/benchmarks/babaisbench/
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.
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.
Not many people share setup with actual setup handholding so that was very G of you
https://x.com/mkagenius/status/2093730391429685732
(xcancel seems to have received a cease and desist)
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.
[0]https://github.com/antirez/ds4
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.
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.
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.
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.
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...
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.
---
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'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.
Yes but it's easy to replace them.
The main reason should be privacy.