Ashen | AI Guy
Region: US
Sunday 21 June 2026 15:13:10 GMT
13989
555
57
53
Music
Download
Comments
Daniel🇧🇬 :
in 5 years when the current hardware right now becomes cheap, we’ll actually have good local ai accessible for everyone
2026-06-22 04:59:11
19
jonny ray :
Nvidia just announced 768 GB RAM machines
2026-06-22 07:19:14
4
bitch_peas :
Can you daisy chain them together? I thought you had to purchase a single piece of hardware to run in?
2026-06-25 03:31:45
1
capet :
Yeah but at what gig to maintain the speed? We are all used 2. With typical api speed
2026-06-22 07:29:00
2
User32456 :
Yeah u need $5k pc to run that
2026-06-21 15:28:42
9
Hammy :
$20k today to run LLMs tomorrow
2026-06-23 02:30:34
2
edo :
ok
2026-06-22 23:03:45
2
bartabee :
the 200gb version is the 2-bit quantize, its only around 80% as accurate. it'll feel pretty dumb for any serious task. the best bang for buck quantize is dynamic 4-bit but youre looking at 500-600GB of RAM just to run it.
2026-06-21 15:56:40
18
Xavier :
man I wish I bought the stuff I wanted before prices spiked
2026-06-24 03:15:12
1
joerhartman :
I think the non quantized version (FP16) requires about 1.6TB of RAM to run at the bare minimum comfortably. Not even 6, 512GB mac mini’s can run this beast (which makes sense given it’s a 753B parameter model). A breakthrough for open source models for sure, but no one is going to be running 5.2 at frontier level performance locally without some serious capital. Hopefully we see developments in the memory space that allow for more memory to be produced cheaply
2026-06-22 08:00:21
2
sah :
that’s just for the quantized version too
2026-06-21 21:03:39
3
aaronanderson23510 :
How are you running 250gb of ram in your Mac mini? I feel like I'm missing something
2026-06-23 20:52:22
2
**** :
Glm5.2 outperforms fable5 on some benchmarks
2026-06-22 21:10:28
3
Micheal :
Finally someone talking facts not trying sell us some crap. Thank you.
2026-06-22 05:53:29
2
Sana Style :
💐
2026-06-22 06:07:28
2
To see more videos from user @midirbot, please go to the Tikwm
homepage.
