Last Notes
Non flash weights just dropped
https://huggingface.co/zai-org/GLM-5.3
Blackwell pros are almost $20k and that's before you put it into a $5k rig.
Gb10 bandwidth scales the more units you add with tensor parallelism
Mac studio for sure, if it will even be possible to buy, which it probably won't
This is running nvfp4, you have to run a patched backend with unmerged PRs to get it working
Speed demo: Glm5.3 flash on 2x GB10s
https://relay.utxo.one/8125e248f9192acc5eb8cfcb812e6708c4eefdcedde6bdb1f885626180ecd779.mp4
Only on real hardware not my Lego PCs 😭
I am hearing that china is about to dump so much ddr5 on the market, cxmc just completed some factory waffer thing, possible cope post, idk, ram prices are whack son
Gonna try omarchy, popos been annoying AF since going to cosmic
Probably gonna waste 4-5 hours trying to get 4-5 more Tok/sec today, thank God AI is here to save me so much time
Then as soon as it's optimized a new model will drop and the cycle with repeat
Glm5.3 makes it official, only clanking sovereignly from now on
https://relay.utxo.one/935dad8920ce8a043e0fbee3a329bcd4799715e994026a3ad146ebaab53fef2e.jpg
Harvey boy enjoying a chicken carcass
Dogs can safely eat chicken bones if they are raw
https://relay.utxo.one/0db37bd7f929055c5acc33ad5c7bd859de129217fc90ab4a73d71683ab80aeeb.mp4
Saylor posted an AI video of himself having sex with a bull
https://relay.utxo.one/8d0668a9fdea0324d14c3b33d140fc153901f3ba5a1c96d2797a6224d5ea88b7.jpg
Trinidad curry #foodstr
https://relay.utxo.one/8124fc8da054614bc54280c8ed62501f15d0831aa98435ed83da5fda7588b2ee.jpg
Same. It's sick. Kinda slow tho. Wen dspark
New release should fix it, out now
🦜🏴☠️
https://relay.utxo.one/4758fd5680783f9d667772f8c65a7b720e90392fa3242cc85c928f2867a32e92.jpg
Imagine how much bandwidth they use
For sun protection I use a hat and shade
Tool calling currently bugged out on it waiting for a patch
My skin care routine is don't put any products on my face
I am running models on 3090s...
It's a fun psyop to fall for speaking from experience
Here's an example of what GLM5.3 Flash produced with one prompt locally, all I told it was to make a reimagined nostr UX using the outbox model.
https://bloom.utxo.one/
Don't get attached to it, I will likely take it down at some point
Nvidia looking to buy hugging face for $13 billion:wisp_eyes:
That hasn't been true for at least 4 years now
Good morning,
Glm 5.3 flash is no joke an approaching endgame level intelligence that you can realistically run at home, especially with these new Mac studios
Re: cost, think of this
If a landscaping company bought $10k of equipment for this business, you'd consider that completely normal. And so a dev, who probably makes more than a landscaper, can justify having a 10k computer for their work.
(I hope you enjoy my cope excuse and go buy some compute)
GLM absolutely dominating DeepSeek on quality, this is sick
Currently getting about 30 Tok/sec decode on GLM flash. Definitely not as fast as deepseek (today)
When dspark comes out this will maybe go to 40-45
Glm is definitely trained on our notes, it did this fully offline
https://relay.utxo.one/20f3039425f0c1f971abe4a0e3d35d17397c29d55bd1ef2750959499d0f5f51e.jpg
Blowing my brains out trying to get glm going locally but we are close gang
:wisp_eyes:
#nevent1q…sz5f
They weren’t hot they were the leftovers
Hmm I wonder if wisp isn't sending the client tag on these
Why arrested? Right to bear arms
Running qwen3.8 flash :wisp_salute:
Lol bitcoin
#nevent1q…r88s
