Continue watching?
Resume from 00:00
00:00 / 00:00
Sleep timer ended
⏪ 10s
⏩ 10s
94.9K views
2d ago
Start building your website today! Visit https://squarespace.syuh.net/LTT and use offer code LTT for 50% off Core, Plus, and Advanced annual and monthly plans f…
...more
100 Comments
Its easy to blame Scam Altman when you have a working relationship with Nvidia
I could have bought the 256 GB DDR5 RAM last year before the price surge, but I was hoping for a better deal on Black Friday.<a href="UCkszU2WH9gy1mb0dV-11UJg/b8IfY7zOK9iVkNAP_I2A-AY"></a>
THAT POWWWWWWER SUPPLY!!! 😍
<a href="https://www.youtube.com/watch?v=vDmjw-c_R2k&t=1709">28:29</a> stfu, just ask your team to research first! Run qwen3.8Flash Next use strata, not an outdated lmstudio. That setup is fast (50-60tps) even on a 4070. TTFT is less than 1 second.
The bigger F-U to Sam would have been building a gaming rig with as much RAM as you possibly can and then letting it idle in a closet for 20 years.
I like your shirt
I'll put on my tinfoil hat and say that it's pretty convenient that as soon as personal use AI is becoming a thing, the gpu/memory "crisis" ensures the average Joe can't possibly do it privately and everyone HAS to buy into a subscription model to make it useful. Almost like they saw this coming.
If I had the cash, I would actually buy this case for 1 reason and 1 reason ONLY: IT HAS WHEELS. Meaning I can put it on my carpet! Not really LMAO
Give me the Ai dude
Perfectly timed!<br>I do believe you planned it for today - but rushing this for today is just perfect
Integrate it with your unifi Stuff over Unify MCP - works great for me
That case is almost as big as Linus 😂 That is a nice piece of hardware mind (the case).
Waited a few years when the PC marketplace was terrible from 2020 to early 2023, bought in early 2023 when stuff wasn't too bad and since then all my stuff has only appreciated. Eventually it will drop down again, just save money now for it and buy super high end and rock your system for like 10 years.
if you want a fast model inference speed, use MOE models by Qwen, like the Qwen 3.6 35B A3B.
I find amazing how their ideologies are getting on the way of the content.
of all the things your workplace has that mine doesn't, I think it's the lighting in that room that I covet the most
"8 inch pounds" Come on guy its the end of 2026
3000w is crazy indeed
Ive been running openclaw on my old gaming rig to much success! building a Standalone system now with nvidia tesla p40s now to see if i cant make things better. Going to try LM studio because ive just been using openclaw but hermes and lm studio looks nice too!
Wow, a TJ07.... I still have mine sitting in my old tech closet. Mine is a custom version that I had fully powder coated way back when. But someday I plan to try and find the hdd cages and a front bracket that would let me fill the front with two or three fans and then turn it into my new server case. And yes my server does deserve to have a window. lol
<a href="https://www.youtube.com/watch?v=vDmjw-c_R2k&t=1565">26:05</a> Incredible height 😮
Ram was promised to him 3000y ago <a href="UCkszU2WH9gy1mb0dV-11UJg/W8IfY_bwAfiPq7IPvNCA2AU"></a>
Sam doesn't think I should have RAM at home. I don't think he should ram his sister.
The timing of this video in this climate is a peice of art.
Hermes! ❤
This video with Emily Young would be better.<br>A server with gui makes no sense.<br>Hermes web is a container why you installed the graphical interface?<br>Why your setup is not rackable?<br>This should be a Linux video.
My 32GB RAM says hi
>"I want it to be private!"<br>>"Also, I'm connecting it to Discord."<br><br>Fucking LMAO, Linus.
i just want prices to go back thats all
Sooo $16k worth of gpus. At even the $100/month that's 160 months or 13 years worth of the subscription. I guess you can never resell the subscription for sure but man, the economics of this aren't straight forward lol. And I'm not even mentioning the RAM...
I bought 16 gig ram and was going to wait for the price to drop it went up 100 dollars
haircut for xmas 😾💯
Could you have the labs explain how they are getting 3d modeling working on such a "small" model? I can't get it working well without using opus 5.5 or gpt 6. Even if I give it the dimensions and hook it up to auto desk fusion which its unable to make anything useful.
I might not be able to afford this or the power bill that it would run, but I can be happy that the plugs at my house are 230V, so I <i>could</i> eventually get a baller 3000W PSU
I got two DGX Sparks and shard models on them. I recommend OpenWebUI because its great. I also got a 5090 FE. I got all of that at the beginning of the year in a panic that everything would get worse and now both are up by multiple thousands. Now I am poor in money but rich in toys.
Holy shit nobody in this video has a clue how to run a local model properly. Wrong model, wrong engine, wrong settings, wrong version hermes.<br><br>Qwen3,8 flash next running on strata engine with hermes desktop and not cli is a much more intelligent, faster and easier to use setup. You would easily get 200 tps with this setup.
This was relevant about a month or so ago. Now you can run the 125B model Qwen 3.8 Flash Next on a 5070TI 16GB VRAM, 64GB RAM via Strata (a llama.cpp based inference server optimized for MOE models) and get between 70 to 110 token/s.
2x m5 ultra linked through thunderbolt 5 would be better build than this.
in 6 months no one will be able to put together a new computer.
Mr. Linus, I enjoy content, but why do you have this disregard and this don't care attitude towards products, like for an example almost in every build video, you show carelessness, you dropped the power supply, sometimes, you just throw stuff like a motherboard in the case, wrestle with stuff that is delicate, and overall very clumsy attitude around hardware in general, idk it bothers me a lot, and i am not the one to nit pick, but it seems like you need to show, ouh i get so much hardware for free, i don't give a shit
2x5090 to run Qwen3.8-27B🤣😂
You joked about water and electricity between one day and the next, but I would genuinely love to see how whether people running local LLMs would be more environmentally friendly than data centres
Is that 3000w psu for the rest of the world that uses 220-240V? Only USA, Canada, Mexico and Japan do not use 220-240V.
<a href="https://www.youtube.com/watch?v=vDmjw-c_R2k&t=515">8:35</a>, this is how my AM4 CPU is anyway, at least when I take try to take the cooler off.
So I took them!... 🤣
Using that qwen model was certainly a choice... I use Gemma 4 26b which is an MOE model on my mac with hermes and it seems to run really well, even without two massive gpus
64GB of VRAM and you picked Qwen3.8-27B?????<br><br>You can go WAY larger with that much VRAM and gobs of RAM, wtf. 16GB VRAM and 64GB RAM is Qwen3.8 Flash Next (175B-A6B if you include the engram table) tier these days. I think it's crazy to not be trying something like GLM 5.3 Flash on Project Maya or Deepseek v4(.1) Flash on Strata/Basalt.
I work at a Amazon Warehouse we have 2 rtx 5090 and i think only the Box is damaged we throing These away my Heart Hurts greetings from Germany
For people who still think LTT builds aren’t janky, show them these clips <a href="https://www.youtube.com/watch?v=vDmjw-c_R2k&t=1211">20:11</a> and <a href="https://www.youtube.com/watch?v=vDmjw-c_R2k&t=1244">20:44</a>
"On my hardware" Yeah let me just get my spare 5090's out. My homelab is running on an 11th gen Intel laptop and an old Datto backup device with a Xeon in it. I am not saying this video isn't cool, it is, rather I wish I could run models locally.
My jailbroken PS5 runs an LLM and Stable Diffusion.<br>Eat my ass, Sam.
wow i felt his disappointment at the end
Linus, what do i have to do to get just 32GB DDR4 Ram and 1TB SSD from you??? please gift it to me. I will do anything
We love you for thisb
That case going to run like a y40
It's hard to actually make local AI safe. Daniel jindo and KaiexplainsYT, both have important videos on a bunch of other stuff you have to do to actually run local AI and have it be private and not leak data all over the place.
You can still buy a Blackwell 4500 pro with 32 gig of RAM for $5,000. It's still a lot and around a 5080 0 in performance, but it's way more affordable for 32 gig of RAM than a 5090
Finally a build without the poor mistake Epic, that is also aged like rotten milk.
Nice video but the lighting is poor, difficult to see things properly when you're working on a black case.
Missing in this video: price breakdown. I do understand the privacy angle, but the subsidized GLM sub is still unbeatable in value.<br><br>At the same time - the best hardware to run local LLMs right now is the one that you already have - you'd be surprised how much value you can get out of your existing machine without spending anything on top.
This video gives me “Linus builds a computer” vibes and I’m all for it
at <a href="https://www.youtube.com/watch?v=vDmjw-c_R2k&t=210">3:30</a> you can see and hear Linus creaming in his mid-life crisis y-fronts.
Ctrl-Altman-Delete. Push a geek, enthusiast et al back to extremes with money (Large Numbers!) and they revert back to nerd...not known for being the best socially and now they're on the world stage... Get fscked literally...it worked before. Mutually Insured Cash is the only reason they're in the same room. Look how uncomfortable they get when the preppy guy hinging popularity instead of good character shows up chasing social status and you've got influence.
<a href="https://www.youtube.com/watch?v=vDmjw-c_R2k&t=1080">18:00</a> Been in this situation many times. Engineer trying to give simple answer to simple question. But the knowledge of how things work, brings up this speech pattern that can be viewed as uncertainty. So from this exchange i can tell Lucas knows quite a lot about power supplies.
Jarvis?! I named mine Holly. It's a nod to the UK series Red Dwarf.
Rich Guy with literal PC building business dunks on Rich Guy is kinda tone deaf though.. this video isn't helping the hardware crisis?
Concerned about privacy: connects his LLM to discord...
Unsure how they at 10tk/sec , I run a bigger version of the same model on a single 5090 and get 10 times that speed , local AI is not that slow , kind of missing the point
need a stupid 240V circuit for 3000w power supply
Nice !
Well, that's great for someone who can afford it........ You you even understand who your viewers are anymore?
Peak three video run we are so back
case looks like a dumpster lol
Why are you not running Canadian LLM?
First LTT video in a long time to catch my interest. More home LLM stuff
<a href="https://www.youtube.com/watch?v=vDmjw-c_R2k&t=7">00:07</a> what are those speakers
<a href="https://www.youtube.com/watch?v=vDmjw-c_R2k&t=1370">22:50</a> Linus looks like he was sitting on a chair when he next to this guy. My god, this dude is tall as fk 😂😂
The title already deserve a like
<a href="https://www.youtube.com/watch?v=vDmjw-c_R2k&t=707">11:47</a> i wonder if those rubber grommets you reversed could have been moved over to the other holes your cables now go through..hmmmmm
Hermes is awesome! I live somewhere we have daily rolling blackouts, So I plugged Hermes into my batteries. It keeps them charged, detects the battery level, adjusts my PC performance accordingly, and, if I'm outside the house, it lets the family know! That the powers out so we can plan accordingly.
Still patiently waiting for Sapphire Rapids to discount enough so I can replace my third-hand Extreme system… 😂😂😂
Can't he download some ram ?
I've been waiting for years to do this, but the more I think about it, the less worthwhile it seems. The upfront cost is ridiculous, the performance is underwhelming, and even the electricity bill alone can end up being higher than the monthly subscription to a frontier AI model.<br><br>I get the whole privacy argument and all that, but unless there's a major breakthrough in efficiency, I really don't see any compelling reason for home users to go down this route.
<a href="https://www.youtube.com/watch?v=vDmjw-c_R2k&t=400">6:40</a> Linus AI confirmed?
Use qwen3.8 flash next, its way faster and also smarter than 27B
is qwen 3.8 27B the actual model you are using? can you not do much much better with that hardware? I can run that on a MacBook Pro. also on first load, it was probably downloading the model, that's why it seemed slow?
A Linus Tech Tip! Loved this!<br><br>So this is 30K build... I guess it's misconfigured, because it should be LOTS faster, the 5090 has insane bandwidth at 1.7TB/s. Quick calculation gives a 70TPS for that model.<br><br>Everyone is sleeping on AMD. The 7900XTX has 24GB of VRAM at decent bandwidth, for the price of half a 5090 you can fill that case up and have much more vram.<br><br>For LLMs AMD actually works painlessly. Not because ROCm works well, it doesn't. But because Vulkan does and llama.cpp targets vulkan. give it a try.<br><br>27B dense is a decent fit for 32GB cards. If you can split it up I would go higher parameter and lower quantization. a 70B or even a 120B model could fit, and the harness would work better, and probably not be much slower.
For those yellow xt power connectors, a general rule is "round is ground" <br>The rounded end of the connector is always ground, the squared end is always voltage.<br>We use xt60 and xt30 connectors on FPV drones and they work great
you wasted so much time building a system that can run tiny models super slow when you could have just gone with the sparks you already got setup there. even just one spark would outperform that system at the model size you used on your build.
LMStudio is LOL
<a href="https://www.youtube.com/watch?v=vDmjw-c_R2k&t=184">3:04</a> poor dude went sky diving 🥀
Bro really used qwen 3.8 27b on 2 5090s 💀
Goddamn i was wondering when you linus going to talk about thiss, I AM SICKKKKK OF FUCKING CHATGPT TOKEN LIMIT BULLSHIT, i paid enough money to them.<br><br>But just a month ago i installed Coucou and Ollama, what i did is im trying to build my own local AI that can work with internet connection and without internet connection, which is a huge massive plus for me, cause what it needs is just my gpu vram and other stuff on my pc, FOR FREE!!!! But still needs to invest on a gpu that has much more vram.
Is that screen from PT a common screen from some NAS or something I don't understand, or did the editors just pull a screen from PT?
fuck ai but cool setup
Imagine if AI gets advanced to the point where it's able to decompile and optimize itself so suddenly all of the accumilated VRAM that AI companies have been hoarding becomes worthless, because anyone will be able to run it on their home computer. That will be a beautiful day
If Labs have so many AI benches assembled, why aren't we getting more AI tinkering content?
I don't really understand the use cases for local agentic ai, particularly for home assistant. But I do love the idea of easily messaging the agent.
I hope Sam Altman will donate 95% of his wealth to charity, because he only uses 5% of it, which is a terrible waste.
my gosh is not the ceo anymore so he can't get sued for the content cheif creator title advice
Up next
From this channel
