I play with local LLMs a lot. I've spent more on hardware than I should. I'm friends with a local group of people who have spent a lot more than I have.
The warning I would have for everyone is to temper your expectations and read the fine print carefully. The big build in article starts off with a $40K budget and then includes 4 GPUs that are $12K each. For those doing the math, this build is going to cost more like 50-55K.
Local setups also often rely on quantization and techniques like REAP to fit the models on their hardware. You will read a lot of claims that 4-bit quantization is lossless, but those claims come from KL divergence measurements on a small corpus. Use one of these 4-bit models on long context coding tasks and the quality will be noticeably less. Even for non-coding tasks like dataset analysis, I can measure a substantial quality difference between 4-bit models, 8-bit quants, and even some times the full 16-bit source.
This article is also encouraging the use of a REAP model, which means someone has cut out some of the weights to make it smaller. The idea is to remove weights that are less useful for certain tasks, but again this is going to reduce the overall quality of the output.
The trap is that people say "I'm running GLM-5.2 locally!" and it sounds amazing when you look at the GLM-5.2 benchmarks. However they're not actually running GLM-5.2, they're running a model derived from GLM-5.2 that discards most of the bits and drops some of the experts. It does not perform the same as what you see in the benchmarks. In my experience, the divergence between a quantized/REAP model and the parent model is unnoticeable when you try it on very small tasks or chat, but becomes painful when you start trying to use it on long-horizon tasks where little errors start compounding.
Then you get into the slippery slope of thinking you're $50K deep into this project, but what you really need is just one or two more of those $12K GPUs to use the next level of quantization that might improve the quality a little more and make your investment worthwhile...
I would very much recommend first using a cloud vendor and setting up an LLM running on there to get a taste of what it’s like before buying the full hardware.
This is similar to my experience with (8-bit quantized, non-MOE, 26b) Qwen locally on my computer. It’s really good for small tasks, but the first time I tried to do a major task with it it straight up forgot what agent harness it was in and started using the wrong format for tool calls lol
(If you’re curious, it was running in Pi, but somehow convinced itself it was running in Claude instead and started trying to call Claude tools that didn’t exist)
I’ve found ds4 on my mbp to be very useful, bought before ram prices became insane. It’s not writing entire applications on it’s own, it has resolved annoying networking issues on my tailnet that I had neither the time nor inclination to figure out on my own and I often find myself reaching for it for simple but annoyingly research intensive tasks that I wouldn’t have otherwise gotten to. Is it opus? No, but is it useful? absolutely and I don’t have to worry about whether or not I’m getting value out of a subscription or the api cost of using it.
> The warning I would have for everyone is to temper your expectations and read the fine print carefully. The big build in article starts off with a $40K budget and then includes 4 GPUs that are $12K each. For those doing the math, this build is going to cost more like 50-55K.
> Local setups also often rely on quantization and techniques like REAP to fit the models on their hardware.
This seems to ignore the very real possibility of running SOTA models at full precision on ordinary local hardware using SSD offload. Yes this will be slow and usually have very low throughput (even batched decode can only achieve so much before power and thermal limits become important, and that still leaves you with slow prefill as a major bottleneck) but that's OK if you aren't expecting a real-time response to begin with and your volumes as a single user are low enough.
SSD streaming throughput is too slow to be usable.
GLM-5.2 has 40B active parameters at a time. At Q4 that's 20GB. The best PCIe 5 SSDs can get 15GB/sec when everything goes well. Every expert load would take more than a second.
If you had enough RAM and enough SSDs in parallel you might get a couple tokens per second on a good day. If you left this machine running 24 hours straight, you might be able to get 200,000 tokens generated.
So it can be done, but only if you interact with your LLM like you're e-mailing someone back and forth and you're okay waiting until tomorrow for a response.
You would spend $50K to buy a machine that consumes 2000W and takes all day to produce as many tokens as I could buy on OpenRouter for $0.60. You would spend $5-15 on electricity depending on where you live.
If you have no other option but to process data locally and you must use a very large model and you aren't in a rush, this can do it. I would not recommend it unless you're desperate and operating inside of rigid constraints.
You can improve that with speculative preload. I'm sure models could be designed and tuned around efficient SSD offloading to keep throughput pretty high.
Wonder if AMD MI350P release will affect setups like this. From what I've heard, the price will be pretty similar to RTX PRO 6000 while having 50% more VRAM which is additionally an HBM3E instead of GDDR7.
> The big build in article starts off with a $40K budget and then includes 4 GPUs that are $12K each. For those doing the math, this build is going to cost more like 50-55K.
Just two months ago you could get RTX PRO 6000's for about $8500 on ebay, which is the MSRP.
All very true. Right now, running GLM 5.2 at its full BF16 quantization level needs 1.5 TB of VRAM. You can't run this locally at a usable speed for less than $250K or so, and frankly I'd be surprised if it could be done for less than $500K.
The best NV4FP quant for 5.2 appears to be lukealonso's at https://huggingface.co/lukealonso/GLM-5.2-NVFP4, and it is capable of good throughput (75-100 tps) without losing much reasoning performance. Allowing for overhead for the KV cache and other requirements, this quant will (barely) run in 8-way tensor-parallel mode on 8x RTX 6000 cards. Not too long ago it was possible to put an 8x machine together for less than $100K USD, but that's probably not true now, assuming you buy all-new components.
It'll almost certainly be worth it, given the abusive behavior we've seen and will continue to see from the major closed-model providers. If I hadn't already put a similar rig together, I'd be kicking myself. But getting it running well is by no means as simple as buying a bunch of RTX6K cards and calling it a day, and people need to know what they're getting into.
Local AI is in its Altair and IMSAI days. There's no turnkey Apple II or C64 on the market yet, much less an IBM PC. Hardware, yes -- you can buy a capable box off the shelf from various vendors -- but you have to be prepared to take up a whole new hobby when it comes to getting a complete system working well.
> hedge against the various tail risks of third-party providers raising prices
They could 10X the prices and you’d still be better off. It’s also unlikely that prices go up enough to warrant a $100K local investment to prevent paying a couple bucks per million tokens.
> or denying you service
I guess you’re not familiar with OpenRouter? There are many providers there. There are providers outside of OpenRouter. There will always be someone to take your business.
> or somehow abusing your data...
If data security is your concern then you’re better renting a server as needed still.
If you cannot tolerate any data leaving, then local models are the only way. You pay a high premium for it!
Raising prices is not a tail risk, anything a local LLM setup can do for you can be done by any cloud provider, with the same capex as yours (or less), there is no moat here, so it is highy price competitive and will remain so. If you want to speculate on hardware shortages, that is a different business altogether and you need no janky garage setup to profit.
Also agreed, it's definitely a sucker's game to run a high-end model locally, by any objective measure.
Still... if it's not your weights, running on your box, you're always going to be behind somebody else's 8-ball. Everybody has to decide for themselves where their priorities lie.
Everything in this post is spot on and it is a rare example of a HN person not saying BS about LLMs!
That said, modern LLM sampling algorithms like min_p, top_n sigma , etc heavily mitigate the performance penalty you get from doing long context tasks. Problems with long context come from accumulation of small sampling errors over time.
My qwen 3.6 27b (the dense one) runs perfectly well on coding tasks at the edge of its context window because I run it using modern LLM sampling stack, namely top N sigma of one, using DRY to stop repetitions and XTC as a superior alternative to temperature for diversification.
Yes there will be a paper soon on arxiv and hopefully NeurIPS proceedings talking about this phenomenon because it’s not well appreciated by the academic AI community yet.
Can you please share you llama.cpp server parameters to turn on modern LLM sampling stack?
Docs [1] say that the top_n_sigma is already in the default sampler list:
"(default: penalties;dry;top_n_sigma;top_k;typ_p;top_p;min_p;xtc;temperature)"
There also exists an in-between possibility, that is, if you get 128GB of vram (there are now multiple options in the market to get that amount with a unified memory architecture) you can run DeepSeek V4 flash at good speed via DwarfStar. I'm not going to spend money on this, but my gut feeling is that this would be the right compromise for a lot of people.
I just started using it on an m4 max 128 and it's the first time since buying the machine a year ago that it feels like local llm "just works" for reasonably decent coding.
Use pi though; claude code has way too much bootstrap context; slows everything way down.
> "~$40k At this price level, you get the next step up in model intelligence. Something pretty close to Claude Opus."
That is equivalent to 16.8 years of Claude Opus 4.8 or Codex GPT 5.5 at $200/mo.
I'm a huge fan of running local models, but they're still wildly expensive, lower quality, and possibly dangerous (if backdoored). I sincerely wish this wasn't the case.
That $200/month is already more like $4,000/month if you have to pay full API pricing - "enterprise" companies for example. That drops the equivalent to 10 months.
(I'd be surprised if that local rig really can drive the equivalent of $4,000/month of API spend though, given that a local rig can run prompts in parallel a lot less effectively than Anthropic's many data centers.)
I agree with your point, but it should be noted that this assumes consistent prices for LLMs. The OpenAIs and Anthropics of this world are still selling the plans at a subsidised prices with the power of VCs, who are going to want that return some time.
You can use a lot more tokens on hardware than you can spend on a $200/m plan.
Inwrnt through 1B tokens my first month with an OEM spark. That's more than $1k of opus. Not a fair comparison, because token patterns are different, but since that time I have also seen a 2-3x improvement in then speeds.from improvements in vllm (mainly MTP). DiffusionGemma is around 4x regular gemma.
You don't own your fiber connection. So why try to own another rapidly depreciating, expensive, and annoying asset?
Rent cloud GPUs!
You get to participate in the ownership, data control, price control, and hacking culture without having to Frankenstein some hobbyist box that costs a ton, is distilled down to functional uselessness, and is a PITA to maintain.
If I'm gonna rent cloud GPUs I might as well just use a subsidized cloud agent like Claude or Codex. As for depreciation, that is true, but the bet is that models get better for a certain parameter count faster than your hardware becomes obsolete, such as Gemma models for example at the same 30 billion parameter count being much better than some years ago.
"A great way to go is 2x RTX 3090s for a total of 48GB VRAM total. You can then run Qwen3.6-27B, which is an awesome model."
Just want to note that for $3k you can get an M5 macbook pro with 48gb of shared memory, and it will not be a giant box. Also, consider committing to spending that money on a cloud hosting provider, which will be at least somewhat cheaper if not significantly cheaper. It is awesome being able to run models locally though.
The cool thing about the 3090s is the RAM bandwidth. Token generation is mostly bottlenecked on memory bandwidth. Dual 3090s have 1.87 TB/s memory bandwidth (0.936 TB/s each), vs the M5 Macbook pro with only 0.3 TB/s (max chip has up to 0.63 TB/s but it's a $10k machine at that config).
This translates to qwen 27b actually working fast enough for useful work on dual 3090s and being painfully slow on Macbook Pros. Also if you're running a big model on a macbook pro the UI gets laggy and the keyboard gets hot. Much better to run dual 3090s in your basement and connect to them from your Macbook.
You need the 128gb ram config to get the 614 GB/s bandwidth (which is $6999), you could skip out on upgrading the storage to save money but at that point I think most people upgrade the storage too at which point it's $8-10k + tax.
No? Any M5 Max with the upgraded GPU has the full bandwidth, which includes the 48GB model the original poster mentioned. Same as the M4 Max, where only the trimmed part had a lower bandwidth.
Why are you throwing in extra cost for something thats not necessary? I know multiple people with 128GB Macs and none of us upgraded the storage. Especially not on a Studio (which isn't currently available).
I will say that their $3k number is off. I somehow missed that, and its too low.
I made a mistake, there is a $5k config with high memory bandwidth. The Max chip has two tiers (I incorrectly thought the tiers were based on memory capacity), you need the higher tier Max GPU upgrade (+$300) to get the 614 GB/s memory bandwidth but you don't need to upgrade the RAM to 128 GB to get the full memory bandwidth. So to get the 614 GB/s you need to upgrade to the max chip + upgraded GPU, but you can spec it at only 48gb if you want. So the total for an m5 max with 614 GB/s memory bandwidth is $4999-$9999 depending on config.
Still 3x lower memory Bandwidth than a dual 3090 setup which you can build for $3k with parts from facebook marketplace and run in your basement.
I have an M5 MacBook Pro and I also have a separate GPU setup for running models. The difference in speed is significant. It's not just token generation speed, but time to first token (prompt processing).
The M5 hardware is amazing for what it is, but GPUs are still so much faster.
Running the models on the GPU box also means I can use the laptop on my lap instead of turning it into a hot plate.
Yeah but 4 bits very often loops needlessly. Which is not that bad because you do not pay for tokens. But you paid for hardware and you want use it for something useful. Q6 is better but then you have like 40t/s prefill. Which is really tiring. But at least it says sorry when you ask it what is wrong! I heard there is some extension for PI preventing that. I need to look into it.
Otherwise I am quite happy.
"Very often" sounds like a lot more than I would say. I've been using Qwen 3.6 27b Q4 in Pi (with out any anti-looping extension) daily for weeks now, and I've had it get stuck in an infinite loop maybe 3 or 4 times.
That's a reasonable option, just be aware that you get about 1/3 as much memory bandwidth with the M5 Pro, or 2/3 with the M5 Max [now you're at $4100 for the lowest-end]. So both your prefill (flops-bound, M5 has a lot less) and decode (bw-bound) will be slower.
To summarize a video I saw recently [0] rebutting your arguments, MacBooks can get crazy slow when running local models or even just Claude Code and Codex due to their poor implementation, to the point that the laptop itself becomes unusable.
There are other arguments for running an ssh-able box in a closet somewhere too as with KVMs you can give an agent remote control over the machine itself such that it has vastly more capabilities than if it were controlling its own machine it's running on, as well as not needing to keep the MacBook open all the time just to have the agent finish running.
GLM 5.2 is "almost Opus," and it needs at least 8xH200s for comfortable inference (so it's closer to $400k than $40k).
They suggest using this modified model:
>A REAP-pruned (≈22% of experts removed), Int8-mix NVFP4 quantized version of GLM-5.2, ≈594B parameters.
I wonder how it behaves in practice outside of benchmarks. Qwen3.6, even at 6-bit quantization, often gets stuck in loops while reasoning. And here they've also removed some experts. I mean, sometimes an 8-bit or 16-bit small model can be smarter than a lobotomized large model. I heard the consensus is you shouldn't go below 8 bit for coding.
Also, it's not clear what is left of the available context when you try to fit a lobotomized model into 4 RTX 6000s. Anything below 100k is barely usable because it often hits compaction before it's able to gather the necessary context
P.S. found in the repos, 240k context
"GLM 5.2 is "almost Opus," and it needs at least 8xH200s for comfortable inference ..."
What is the behavior if one were to run GLM 5.2 with only a single H200 ?
Would it fail to run at all, or would it just run so slowly as to be unusable ?
I would like to prove out the build, and concept, of a SOTA model locally, but then backfill the rest of the GPUs in 18-24 months when they cost significantly less ...
Say more. My expectation is that the current gen of gpus will start being replaced by the next gen, and then it may be possible to get used ones that are still within their useful life at lower prices. My expectation is also that memory vendors are likely to increase production, which will drive those prices down eventually. Maybe not over the next 18-24 months though.
Looping, like most other phenomenons related to LLMs, is a sampling problem and can be easily solved with the DRY penalty. It’s in llamacpp. The same guy who wrote heretic invented the SOTA antilooping and diversification strategies.
You can get 1M context with the lukealonso NVFP4 quant on 8x RTX6000s, which remains coherent and useful through at least 400k. No real need to run 8x H200s unless you just want to. Or unless you need to serve many concurrent users or agents on a regular basis.
Might as well add my own experience since I just set up a local llm this week. I went with a 32GB card made by Intel called Arc B70, which is cheaper than a 3090 and more has ram, at the cost of a slower memory bus. edited to remove something incorrect, thanks diablod3
I went with this because a) the models I wanted to use are a little too big to fit comfortably in 24gb, plus I need room for a few additional small models for autocomplete and speech recognition, and b) I already had a cheap server to use and dual gpus would've required upgrading the mobo and power supply and probably the case as well.
It was definitely a little tricky to set up. The Intel line requires a driver package called "level zero" to support something called SYCL (Intel's version of CUDA basically, AFAICT) that was tricky to get working. I am running llama.cpp in a docker container, which also required some fiddling to get the container to see the card. You also need a kernel from the last few months.
Once I got it working though, the results are very impressive for a $1k investment. Qwen 3.6 35B at q4 quantization takes about 3/4 of the ram and delivers like 88 tokens/sec. So, if you want a decent-sized model for cheap, this is one way to go.
Whoops thanks, was going from memory. At any rate, the effect is that it's somewhat slower than the 3090, when using a model small enough to fit entirely in nvram, but can fit models the 3090 can't.
Related - what is the best isolation system available? Do I have to go full, fat VMs or can I get by with a Firecracker-like thing?
Seemingly every available option has some subtle-gotchas about how easy it is to blow off your foot and effectively have no security at all. I use VMs because I actually trust that security is a foundational principle of the technology, not a well-if-you-use-these-20-flags-and-squint kind of deal.
It depends - for what? If your security model is sandboxing an agent to ensure they don't nuke your PC, then there are a lot of options, you can use something like bubblewrap[1] or a microVM like libkrun[2] if your goal is light-weight, up to full Docker if you want the tooling that comes with that.
im not sure that there is a plug n play set up that will work for everyone, because as with any security boundary, each layer of hardening has a usability trade-off. i definitely feel you about the uncertainty of it all, how do you actually know everything is tight?
personally, i think either a VM or microVM is the way to go. these things are actually designed as security boundaries, as opposed to containers. and as compared to bubblewrap, you can just give the agent a whole FS to work with and run it in yolo mode, whereas with bubblewrap you have to manually bootstrap the availability of each individual dev tool and make sure its config dirs and package caches and etc are mounted in a secure way and still will probably hit perm errors all the time. and there's just way less isolation.
also, something that has limited support in harnesses but IMO would make a lot of sense is running the harness process in the host, but having all the tool calls and file system interactions delegated to the VM. that way you keep all your session data and auth keys on the main machine where it can never get into context. otoh it makes your harness part of the security boundary, so that's the trade-off.
there's also all the usability questions around how to actually get data in/out of the VM. i have a script which can push local git repos into the VM and then pull from them as a remote, so the VM can't initiate any connection with the host doesn't need to hold git credentials. but ig for someone who wants their agent to push straight to GitHub that's a waste of effort.
options i've tried or seen for the VM itself:
- qemu + libvirt: takes some doing to wrangle it together, but very battle tested and configurable
- crun-vm is a PoC of higher level integration layer between podman and qemu, which is a really cool way to go about it. seems maybe abandoned but i just think it's neat and very existing tools/standards oriented rather than starting a new project and brand so i mention
- libkrun is a newer entrant, and several ppl have built wrappers around it:
- microsandbox
- smolvm (posted/discussed on here recently)
- krunvm
from my understanding, you can run the inference server (llama.cpp/vllm/whatever) and the agent/harness in different contexts, event different machines.
The risky part is in the agent/harness and what tools it has access to.
You don't need to give GPU passthrough to the VM running the agent/harness.
There is still a risk of a prompt messing with the inference server, but I think that's a much lower risk compared to an agent doing whatever on its own.
For qwen3.6-27b you can also run the q4 variant with full ~250K context on one 3090. It's fast enough to not be frustrating so the speed gains with 2x 3090s wouldn't be worth it to me. Running a q6 on 2x 3090s at half the speed with a smaller context is an option, but you're really not going to compete with SOTA models there anyway so unless you already have 2x 3090s, I would say 1 is the best investment given current prices. It's good enough to do a lot, especially with a well-configured harness.
Are you running qwen3.6-27b on one 3090 with your KV cache at q4? Ime there is significant long-context recall accuracy degradation at that precision. I prefer putting the KV cache at q8 and working with the 120k context
Use modern samplers and you don’t need to limit yourself to 8bit at half the context window. I could push it down to 1.58 bits and get decently good output easily by simply not using the garbage default top_p and top_k that vendors continue to wrongly recommend.
No, there are quite a few models which are smaller, more accurate, and faster. For example Parakeet TDT v3 is half the size, way faster, and lower WER. There's also Voxstral, which is much larger but also even more accurate.
But the ecosystem isn't as mature, so Whisper is still a valid option, even now. For example Parakeet uses Nemotron framework (made by Nvdia), normally you need CUDA, so you need to use an ONNX version instead on AMD. Meanwhile Whisper has VLLM and desktop apps like Buzz.
There aren't many benchmarks and they often don't have all the models, since STT doesn't get nearly enough attention as normal LLMs, but this is one of the more complete ones:
https://artificialanalysis.ai/speech-to-text/non-streaming
I don't have anything to compare against, since I have just started using it. But I was fairly happy with it on my personal recordings from my phone. Also, I ran it on my CPU (Core i7) and it was perfectly usable, as something to run when not using the machine for anything else.
There's a sub 2k tier with a single 3090 that's also serviceable. Run https://github.com/noonghunna/club-3090 with beellama, fast inference at the cost of a reduced 102k context window
I picked up the 128gb version when it was $2,199 and it runs Qwen 3.6 reasonably well with a 128kb context. Not very useful for complex tasks but it can handle some web stuff.
I've been happy with an OEM Spark (128G), enough so that I picked up a second one. Have 2x qwen and 1x gemma (both at 8bit and full context), plus embedding, Re-Ranker, and a 1.7B for little things. Running 6x models, probably going to add STT here soon, want to try talking more than typing.
The caveat is that if you try to use multiple models on the same device at the same time, you thrash and destroy tok/s
This is a great guide. However, the economics just do not work in my favor at all. Even if I were to spend $2k, I get much more flexibility of model intelligence and choice from a provider for $20/month.
While I think that local LLMs are the future, i think these setups are insane. You shouldn't be trying to push the SOTA, most people underestimate how much you can get out of small LLMs.
Why ask FABLE 5000 to "summarize this email thread" when a tiny model can do the job?
Sure Codex3000 can oneshot your backlog, but why not use a subsidized subscription to do it for now? We're clearly not at the peak of these model's capabilities yet.
Could someone give me an actual guide for spending as little as possible to get as maximal gains with either SOTA or cheap models as a systems administrator and not someone like a full-stack developer?
The models are so powerful and consequently so expensive and confusing to use, I don't get all of it.
I agree that local LLMs are the likely future and worth investing in… but at $40k for possible-SOTA right now, this isn’t worth it for the average consumer.
I’m pretty bullish that Apple will deliver something very competitive for the average consumer in the next couple years.
Local open weight models will definitely be a future trend. Imagine if an Opus-level model could run locally: many more latent use cases would likely emerge, since Opus is priced so high. Perhaps the future will be a multi-model architecture, where frontier models handle planning and local models carry out the concrete execution.
I recently wrote up how I run local LLMs, because several folks had asked (https://swelljoe.com/post/how-i-run-local-llms/) and I think even my setup, which I spent maybe $4200 on, half on a Strix Halo and half on upgrades for my desktop, would be too expensive to justify today. I bought before prices went through the roof, and only did so because I like to tinker with hardware...not because I expected it to ever pay for itself vs. buying subsidized tokens from the big guys or the cheap tokens from efficient providers like DeepSeek.
Buying four $13000 GPUs and several thousand dollars worth of supporting hardware seems crazy. This supply shortage has to end eventually, and I can buy billions of DeepSeek, MiMo, and GLM tokens, and use $100 or $200 a month subscriptions for the big guys in the meantime for the difference in price once that happens. And, you can't even run the full-sized GLM on that hardware, it is quantized and so is your KV cache; the degradation is small, but not non-existent. You're not running a model that's equal to what you get when you buy GLM tokens from Z.ai.
My recommendation for self-hosting is this: If you already have a 24GB or 32GB GPU, or two, or a recent Mac with 32GB or more, run the appropriate quantization of Qwen 3.6 27B or Gemma 4 31B. If your hardware is older and too slow for that, use the MoE, but know it'll be dumber. Use the tiny model for the stuff that doesn't need deep smarts: Research (give it a Brave or Exa MCP for web search), summarization, simple Python scripts for basic tasks, simple websites or web apps, categorization of stuff (I used Gemma 4 to review my past writing for friendliness and helpfulness), etc. It can also be a sub-agent for bigger agents (for those same kinds of tasks). Gemma 4 12B is an incredibly good model for its size, particularly for vision tasks, and in the 4-bit quantization (7GB on disk) it runs on anything, even a modern tablet or phone.
And, if you don't already have a big GPU or unified memory Mac, just wait. Use the cheap tokens every AI company wants to sell you, for now. A Claude or Codex or Gemini subscription is a good deal. Tokens from DeepSeek are a good deal, especially with Reasonix agent (which maximizes caching, which DeepSeek is uniquely good at, and cached tokens are uniquely cheap at DeepSeek). GLM is Good Enough and has a cheap coding plan. MiMo has the cheapest tokens for a 1T+ model in the game, though DeepSeek and GLM are better models, MiMo is fine.
When prices come down, I'll be speccing out a beast to run the big models, too. But, I'm not paying 4x for RAM and GPU and storage, and y'all shouldn't either. That's crazy. Computer prices go down over time. It is the law.
80 tok/s which is kind of a lot for GLM. My experience running 80 tok/s on other LLM is that it ~seems faster than cloud inference, but that obviously depends what you use, in my case ChatGPT.
What harness is the best for local LLMs? I've been researching optimizing local LLM agent harness performance with context/ tools. Quite the endeavor and would love to learn what users prefer for this type of workflow.
I like vibe and pi. Vibe just looks nice and is good enough. But pi extensibility is just another level. There is also Dirac that is quite OK but seems like full of bugs. Zerostack is the simplest harness I saw. OpenCode is OK too. Rest I did not try.
Apple M series chips deserve a mention as another option, especially since you get a whole Mac laptop or desktop workstation too.
They have unified memory and respectable inference performance, and for some variations can be cheaper than video cards, especially if you get an older-gen high-end M series with a lot of RAM used or refurbished.
I've read that Apple has plans once the RAM bottleneck passes to offer more RAM in all their models, and that future M series GPUs and NPUs will be even better for local inference, so in the future I expect Apple to be a serious offering for local inference and AI research workstations.
And what about AMD and Intel Arc GPUs? They don't get as much love but I've heard they can be compelling for certain shapes of a local LLM configuration.
At this point though, I think we may be in a "renters market" for LLM compute. If you want privacy it might be better to rent GPU time in raw form or use spot pricing at various providers. It probably only makes sense to build if you have extreme privacy/security needs or just want to do it cause it's cool.
Do we have evidence that this will actually happen? Maybe the belief that it won't pass is what requires evidence, but I think there's a widespread feeling right now that things are just getting permanently worse and this is one example.
It'll probably take a few years. There's many fabs under construction.
One thing holding back capacity expansion is that a lot of people are concerned this is a bubble. They're worried it'll pop and leave them with orphaned assets if they over-invest in production.
Of course maybe they're right and that will happen. If the data center construction boom ends, RAM prices will fall.
It can be considered SOTA within is size category. Very useful for many things. You still want access to big models, I recommend OpenCode Go if you want to stay with open models.
The warning I would have for everyone is to temper your expectations and read the fine print carefully. The big build in article starts off with a $40K budget and then includes 4 GPUs that are $12K each. For those doing the math, this build is going to cost more like 50-55K.
Local setups also often rely on quantization and techniques like REAP to fit the models on their hardware. You will read a lot of claims that 4-bit quantization is lossless, but those claims come from KL divergence measurements on a small corpus. Use one of these 4-bit models on long context coding tasks and the quality will be noticeably less. Even for non-coding tasks like dataset analysis, I can measure a substantial quality difference between 4-bit models, 8-bit quants, and even some times the full 16-bit source.
This article is also encouraging the use of a REAP model, which means someone has cut out some of the weights to make it smaller. The idea is to remove weights that are less useful for certain tasks, but again this is going to reduce the overall quality of the output.
The trap is that people say "I'm running GLM-5.2 locally!" and it sounds amazing when you look at the GLM-5.2 benchmarks. However they're not actually running GLM-5.2, they're running a model derived from GLM-5.2 that discards most of the bits and drops some of the experts. It does not perform the same as what you see in the benchmarks. In my experience, the divergence between a quantized/REAP model and the parent model is unnoticeable when you try it on very small tasks or chat, but becomes painful when you start trying to use it on long-horizon tasks where little errors start compounding.
Then you get into the slippery slope of thinking you're $50K deep into this project, but what you really need is just one or two more of those $12K GPUs to use the next level of quantization that might improve the quality a little more and make your investment worthwhile...
(If you’re curious, it was running in Pi, but somehow convinced itself it was running in Claude instead and started trying to call Claude tools that didn’t exist)
> Local setups also often rely on quantization and techniques like REAP to fit the models on their hardware.
This seems to ignore the very real possibility of running SOTA models at full precision on ordinary local hardware using SSD offload. Yes this will be slow and usually have very low throughput (even batched decode can only achieve so much before power and thermal limits become important, and that still leaves you with slow prefill as a major bottleneck) but that's OK if you aren't expecting a real-time response to begin with and your volumes as a single user are low enough.
GLM-5.2 has 40B active parameters at a time. At Q4 that's 20GB. The best PCIe 5 SSDs can get 15GB/sec when everything goes well. Every expert load would take more than a second.
If you had enough RAM and enough SSDs in parallel you might get a couple tokens per second on a good day. If you left this machine running 24 hours straight, you might be able to get 200,000 tokens generated.
So it can be done, but only if you interact with your LLM like you're e-mailing someone back and forth and you're okay waiting until tomorrow for a response.
You would spend $50K to buy a machine that consumes 2000W and takes all day to produce as many tokens as I could buy on OpenRouter for $0.60. You would spend $5-15 on electricity depending on where you live.
If you have no other option but to process data locally and you must use a very large model and you aren't in a rush, this can do it. I would not recommend it unless you're desperate and operating inside of rigid constraints.
Just two months ago you could get RTX PRO 6000's for about $8500 on ebay, which is the MSRP.
The MSRP was raised to $13,250.
Warranty is very important for expensive cards like this. I don't recommend buying on eBay unless they come with a very big discount.
The best NV4FP quant for 5.2 appears to be lukealonso's at https://huggingface.co/lukealonso/GLM-5.2-NVFP4, and it is capable of good throughput (75-100 tps) without losing much reasoning performance. Allowing for overhead for the KV cache and other requirements, this quant will (barely) run in 8-way tensor-parallel mode on 8x RTX 6000 cards. Not too long ago it was possible to put an 8x machine together for less than $100K USD, but that's probably not true now, assuming you buy all-new components.
It'll almost certainly be worth it, given the abusive behavior we've seen and will continue to see from the major closed-model providers. If I hadn't already put a similar rig together, I'd be kicking myself. But getting it running well is by no means as simple as buying a bunch of RTX6K cards and calling it a day, and people need to know what they're getting into.
Local AI is in its Altair and IMSAI days. There's no turnkey Apple II or C64 on the market yet, much less an IBM PC. Hardware, yes -- you can buy a capable box off the shelf from various vendors -- but you have to be prepared to take up a whole new hobby when it comes to getting a complete system working well.
The proper financial comparison for GLM-5.2 would be one of the providers on OpenRouter or renting a server as needed. Compare apples to apples.
You will almost certainly never break even compared to paying per token.
Local LLMs at this scale are only worth it if you have extremely strict requirements that data not leave the premises.
They could 10X the prices and you’d still be better off. It’s also unlikely that prices go up enough to warrant a $100K local investment to prevent paying a couple bucks per million tokens.
> or denying you service
I guess you’re not familiar with OpenRouter? There are many providers there. There are providers outside of OpenRouter. There will always be someone to take your business.
> or somehow abusing your data...
If data security is your concern then you’re better renting a server as needed still.
If you cannot tolerate any data leaving, then local models are the only way. You pay a high premium for it!
Still... if it's not your weights, running on your box, you're always going to be behind somebody else's 8-ball. Everybody has to decide for themselves where their priorities lie.
Obviously depends on the use case and threat model, but that hardware is publicly available at far less than $500k upfront.
That said, modern LLM sampling algorithms like min_p, top_n sigma , etc heavily mitigate the performance penalty you get from doing long context tasks. Problems with long context come from accumulation of small sampling errors over time.
My qwen 3.6 27b (the dense one) runs perfectly well on coding tasks at the edge of its context window because I run it using modern LLM sampling stack, namely top N sigma of one, using DRY to stop repetitions and XTC as a superior alternative to temperature for diversification.
Yes there will be a paper soon on arxiv and hopefully NeurIPS proceedings talking about this phenomenon because it’s not well appreciated by the academic AI community yet.
Docs [1] say that the top_n_sigma is already in the default sampler list: "(default: penalties;dry;top_n_sigma;top_k;typ_p;top_p;min_p;xtc;temperature)"
[1] https://github.com/ggml-org/llama.cpp/blob/master/tools/serv...
Use pi though; claude code has way too much bootstrap context; slows everything way down.
That is equivalent to 16.8 years of Claude Opus 4.8 or Codex GPT 5.5 at $200/mo.
I'm a huge fan of running local models, but they're still wildly expensive, lower quality, and possibly dangerous (if backdoored). I sincerely wish this wasn't the case.
(I'd be surprised if that local rig really can drive the equivalent of $4,000/month of API spend though, given that a local rig can run prompts in parallel a lot less effectively than Anthropic's many data centers.)
Inwrnt through 1B tokens my first month with an OEM spark. That's more than $1k of opus. Not a fair comparison, because token patterns are different, but since that time I have also seen a 2-3x improvement in then speeds.from improvements in vllm (mainly MTP). DiffusionGemma is around 4x regular gemma.
You don't own your fiber connection. So why try to own another rapidly depreciating, expensive, and annoying asset?
Rent cloud GPUs!
You get to participate in the ownership, data control, price control, and hacking culture without having to Frankenstein some hobbyist box that costs a ton, is distilled down to functional uselessness, and is a PITA to maintain.
Just want to note that for $3k you can get an M5 macbook pro with 48gb of shared memory, and it will not be a giant box. Also, consider committing to spending that money on a cloud hosting provider, which will be at least somewhat cheaper if not significantly cheaper. It is awesome being able to run models locally though.
So, I always thought local LLMs were toys not worth pursuing.
Only once have I tried something decent like Gemma 4 31B and Qwen 3.6 27B did I realize how incredibly useful they are.
You stop fearing you are sharing sensitive information.
You stop fearing you will run out of tokens.
You stop fearing about the availability of the remote AI.
Local LLMs are extremely valuable.
This translates to qwen 27b actually working fast enough for useful work on dual 3090s and being painfully slow on Macbook Pros. Also if you're running a big model on a macbook pro the UI gets laggy and the keyboard gets hot. Much better to run dual 3090s in your basement and connect to them from your Macbook.
Even a 128GB is $6.8k today. Still only 2/3 your quote.
Bandwidth is relevant (I have both a 5090 and an M4 Max 128GB Studio, so have direct comparison right here), but quote the cost appropriately!
Why are you throwing in extra cost for something thats not necessary? I know multiple people with 128GB Macs and none of us upgraded the storage. Especially not on a Studio (which isn't currently available).
I will say that their $3k number is off. I somehow missed that, and its too low.
Still 3x lower memory Bandwidth than a dual 3090 setup which you can build for $3k with parts from facebook marketplace and run in your basement.
2x3090 (has an nvlink bridge though it didn't seem to matter hugely for inference)
Qwen 3.6 27b int4: Concurrency 1: 68 tok/s output Concurrency 32: 363 tok/s output Prompt processing speed: 1520 tok/s
Qwen 3.6 35ba3b int4: Concurrency 1: 150 tok/s output Concurrency 32: 1083 tok/s output Prompt processing speed: 4324 tok/s
Macbook Pro m3 36gb RAM: Qwen 3.6 27b int4: Concurrency 1: 18 tok/s output didn't measure the other metrics and it was a slightly different benchmark.
The M5 hardware is amazing for what it is, but GPUs are still so much faster.
Running the models on the GPU box also means I can use the laptop on my lap instead of turning it into a hot plate.
Get a regular laptop and use the network to access the LLM
There are other arguments for running an ssh-able box in a closet somewhere too as with KVMs you can give an agent remote control over the machine itself such that it has vastly more capabilities than if it were controlling its own machine it's running on, as well as not needing to keep the MacBook open all the time just to have the agent finish running.
[0] https://youtu.be/9tGrhrVKCrE
GLM 5.2 is "almost Opus," and it needs at least 8xH200s for comfortable inference (so it's closer to $400k than $40k).
They suggest using this modified model:
>A REAP-pruned (≈22% of experts removed), Int8-mix NVFP4 quantized version of GLM-5.2, ≈594B parameters.
I wonder how it behaves in practice outside of benchmarks. Qwen3.6, even at 6-bit quantization, often gets stuck in loops while reasoning. And here they've also removed some experts. I mean, sometimes an 8-bit or 16-bit small model can be smarter than a lobotomized large model. I heard the consensus is you shouldn't go below 8 bit for coding.
Also, it's not clear what is left of the available context when you try to fit a lobotomized model into 4 RTX 6000s. Anything below 100k is barely usable because it often hits compaction before it's able to gather the necessary context P.S. found in the repos, 240k context
What is the behavior if one were to run GLM 5.2 with only a single H200 ?
Would it fail to run at all, or would it just run so slowly as to be unusable ?
I would like to prove out the build, and concept, of a SOTA model locally, but then backfill the rest of the GPUs in 18-24 months when they cost significantly less ...
going to need you to sit down for this one...
I assume you can then somehow run several hundreds of prompts concurrently?
I went with this because a) the models I wanted to use are a little too big to fit comfortably in 24gb, plus I need room for a few additional small models for autocomplete and speech recognition, and b) I already had a cheap server to use and dual gpus would've required upgrading the mobo and power supply and probably the case as well.
It was definitely a little tricky to set up. The Intel line requires a driver package called "level zero" to support something called SYCL (Intel's version of CUDA basically, AFAICT) that was tricky to get working. I am running llama.cpp in a docker container, which also required some fiddling to get the container to see the card. You also need a kernel from the last few months.
Once I got it working though, the results are very impressive for a $1k investment. Qwen 3.6 35B at q4 quantization takes about 3/4 of the ram and delivers like 88 tokens/sec. So, if you want a decent-sized model for cheap, this is one way to go.
They both have GDDR6.
The B70 has 256 bit it bus at a clock speed of 2375mhz (608 GB/s), the 3090 has a 384 bit bus at a clock speed of 2438mhz (936 GB/s).
It isn't slower, it just has less channels, ie, it is less wide.
Seemingly every available option has some subtle-gotchas about how easy it is to blow off your foot and effectively have no security at all. I use VMs because I actually trust that security is a foundational principle of the technology, not a well-if-you-use-these-20-flags-and-squint kind of deal.
[1] https://github.com/containers/bubblewrap
[2] https://github.com/libkrun/libkrun
personally, i think either a VM or microVM is the way to go. these things are actually designed as security boundaries, as opposed to containers. and as compared to bubblewrap, you can just give the agent a whole FS to work with and run it in yolo mode, whereas with bubblewrap you have to manually bootstrap the availability of each individual dev tool and make sure its config dirs and package caches and etc are mounted in a secure way and still will probably hit perm errors all the time. and there's just way less isolation.
also, something that has limited support in harnesses but IMO would make a lot of sense is running the harness process in the host, but having all the tool calls and file system interactions delegated to the VM. that way you keep all your session data and auth keys on the main machine where it can never get into context. otoh it makes your harness part of the security boundary, so that's the trade-off.
there's also all the usability questions around how to actually get data in/out of the VM. i have a script which can push local git repos into the VM and then pull from them as a remote, so the VM can't initiate any connection with the host doesn't need to hold git credentials. but ig for someone who wants their agent to push straight to GitHub that's a waste of effort.
options i've tried or seen for the VM itself: - qemu + libvirt: takes some doing to wrangle it together, but very battle tested and configurable - crun-vm is a PoC of higher level integration layer between podman and qemu, which is a really cool way to go about it. seems maybe abandoned but i just think it's neat and very existing tools/standards oriented rather than starting a new project and brand so i mention - libkrun is a newer entrant, and several ppl have built wrappers around it: - microsandbox - smolvm (posted/discussed on here recently) - krunvm
this is all Linux oriented, it's all i know.
The risky part is in the agent/harness and what tools it has access to.
You don't need to give GPU passthrough to the VM running the agent/harness.
There is still a risk of a prompt messing with the inference server, but I think that's a much lower risk compared to an agent doing whatever on its own.
This approach requires that you trust the llama.cpp codebase, essentially. It might be reasonable not to.
I suppose in principle there is the risk of a prompt exploit corrupting the inference server.
But the ecosystem isn't as mature, so Whisper is still a valid option, even now. For example Parakeet uses Nemotron framework (made by Nvdia), normally you need CUDA, so you need to use an ONNX version instead on AMD. Meanwhile Whisper has VLLM and desktop apps like Buzz.
There aren't many benchmarks and they often don't have all the models, since STT doesn't get nearly enough attention as normal LLMs, but this is one of the more complete ones: https://artificialanalysis.ai/speech-to-text/non-streaming
I'm curious if GMKtec's EVO-X2, with ~96GB of usable VRAM, is still a good solution for something like this for $3,399.
The caveat is that if you try to use multiple models on the same device at the same time, you thrash and destroy tok/s
Why ask FABLE 5000 to "summarize this email thread" when a tiny model can do the job?
Sure Codex3000 can oneshot your backlog, but why not use a subsidized subscription to do it for now? We're clearly not at the peak of these model's capabilities yet.
The models are so powerful and consequently so expensive and confusing to use, I don't get all of it.
I’m pretty bullish that Apple will deliver something very competitive for the average consumer in the next couple years.
Buying four $13000 GPUs and several thousand dollars worth of supporting hardware seems crazy. This supply shortage has to end eventually, and I can buy billions of DeepSeek, MiMo, and GLM tokens, and use $100 or $200 a month subscriptions for the big guys in the meantime for the difference in price once that happens. And, you can't even run the full-sized GLM on that hardware, it is quantized and so is your KV cache; the degradation is small, but not non-existent. You're not running a model that's equal to what you get when you buy GLM tokens from Z.ai.
My recommendation for self-hosting is this: If you already have a 24GB or 32GB GPU, or two, or a recent Mac with 32GB or more, run the appropriate quantization of Qwen 3.6 27B or Gemma 4 31B. If your hardware is older and too slow for that, use the MoE, but know it'll be dumber. Use the tiny model for the stuff that doesn't need deep smarts: Research (give it a Brave or Exa MCP for web search), summarization, simple Python scripts for basic tasks, simple websites or web apps, categorization of stuff (I used Gemma 4 to review my past writing for friendliness and helpfulness), etc. It can also be a sub-agent for bigger agents (for those same kinds of tasks). Gemma 4 12B is an incredibly good model for its size, particularly for vision tasks, and in the 4-bit quantization (7GB on disk) it runs on anything, even a modern tablet or phone.
And, if you don't already have a big GPU or unified memory Mac, just wait. Use the cheap tokens every AI company wants to sell you, for now. A Claude or Codex or Gemini subscription is a good deal. Tokens from DeepSeek are a good deal, especially with Reasonix agent (which maximizes caching, which DeepSeek is uniquely good at, and cached tokens are uniquely cheap at DeepSeek). GLM is Good Enough and has a cheap coding plan. MiMo has the cheapest tokens for a 1T+ model in the game, though DeepSeek and GLM are better models, MiMo is fine.
When prices come down, I'll be speccing out a beast to run the big models, too. But, I'm not paying 4x for RAM and GPU and storage, and y'all shouldn't either. That's crazy. Computer prices go down over time. It is the law.
They have unified memory and respectable inference performance, and for some variations can be cheaper than video cards, especially if you get an older-gen high-end M series with a lot of RAM used or refurbished.
I've read that Apple has plans once the RAM bottleneck passes to offer more RAM in all their models, and that future M series GPUs and NPUs will be even better for local inference, so in the future I expect Apple to be a serious offering for local inference and AI research workstations.
And what about AMD and Intel Arc GPUs? They don't get as much love but I've heard they can be compelling for certain shapes of a local LLM configuration.
At this point though, I think we may be in a "renters market" for LLM compute. If you want privacy it might be better to rent GPU time in raw form or use spot pricing at various providers. It probably only makes sense to build if you have extreme privacy/security needs or just want to do it cause it's cool.
Do we have evidence that this will actually happen? Maybe the belief that it won't pass is what requires evidence, but I think there's a widespread feeling right now that things are just getting permanently worse and this is one example.
People do that all the time, and sometimes it doesn't pay off.
One thing holding back capacity expansion is that a lot of people are concerned this is a bubble. They're worried it'll pop and leave them with orphaned assets if they over-invest in production.
Of course maybe they're right and that will happen. If the data center construction boom ends, RAM prices will fall.
FUCK EUROPE. FUCK CANADA.
GO HOME JEWS!!! YOU TRIED TO RUN USA - YOU CAN'T DO IT. LOL. RUINED NEWS & ECONOMY. THOUGHT THAT WAS YOUR FUCKING FORTE????? GET OUT.