@Nasheology @ZackKorman The weights release on the 27th. Why would you call BS? What gave you the impression that I am amused by, or have something to gain from lying?
English
mfs
197 posts
mfs
@Mike5tevenson
Systems Engineer. Long time infosec guy, currently into local AI + infosec.
US Katılım Mayıs 2017
134 Takip Edilen24 Takipçiler
@Mike5tevenson @ZackKorman How can you audit Kimi excatly? Reverse engineer the weights? With what compute? I call BS
English
Kimi K3 has 2.8 trillion parameters? Think of all the sleeper agents they can fit in that thing!
English
they distilled our models that we didn't invented yet
Arena.ai@arena
Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 pts, surpassing Claude Fable 5. This is a 17-place jump from Kimi-k2.6 (#18 -> #1). In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics, Consumer Product, Simulations, and Content Creation Tools, landing #2 only in Gaming behind Fable 5. The full model weights will be released by July 27. Congrats to the @Kimi_Moonshot team on this major milestone!
English
@ZackKorman The defender / attacker dichotomy is a big topic, but in a nutshell if we are concerned with vulns, the defender is well positioned. IMHO at that point the issues are attacker scaling and black boxes on the wire.
English
AI doomers want us to believe in the cyber apocalypse, where evil hackers use advanced AI capabilities and cause mass chaos.
There are so many reasons why that is very unlikely.
English
@checkedz @brandonburr1986 @growing_daniel Either parroting or maybe tried running Deepseek on a mac mini? I dunno, but confidently incorrect either way.
English
@brandonburr1986 @growing_daniel Weird flawed reasoning, but either you’re just coping or it’s on purpose
English
@TheZachMueller I meant broadly speaking, if I am interpreting your response correctly.
English
I wish I was making this up. Had a UPS delivery to the house (power cable, he delivered some parts yesterday). We got into the conversation of what's the rig in the house I have going on, into Ollama into Hermes Agent into Qwen3 27B.
This is actually hitting a tipping point
English
@MichaelGannotti Good news for the masses - the moat is constantly being filled.
English
Owning your inference stack isn't cheaper. It's harder. That's the point — the difficulty is the moat.
English
@MichaelGannotti Yep, I need a couple. I’m using this franken-rig right now, but it’s a bit limited for large models. x.com/mike5tevenson/…
mfs@Mike5tevenson
There are many franken-rigs, but this one is mine. It’s extra franken-ey.
English
@Mike5tevenson 😂 The DGX Spark is worth every organ. Honestly the best AI hardware purchase I've made — local inference that actually keeps up.
English
seeing these Kimi K3 posts has me drooling
filipe@filicroval
Kimi K3 also mogs Fable 5 on frontend design! this time it's not even close, Fable 5 didn't even try to produce something half as good as Kimi's result. the prosthetic hand still needs polishing but it's way more capable and Kimi deployed more feature than Fable did. no need to wonder what model to use for frontend anymore, the choice is clear.
English
I fear that this will lead to legislation, across the board.
roon@tszzl
the era of the chinese labs being far behind is over, Kimi is at least on par with the modern public frontier models. people have to think differently now without any competitive margin built in
English
I use local AI, so I’m gambling with my own chips.
kapilansh@kapilansh_twt
vibe coding is just another form of gambling you put money in you hope something works most of the time it doesn't agree ?
English
when you dangle a new AI product in front of n00bs
(I am one, so I am authoritative voice)
The Cats 𝕏@TheCatsX
Total chaos for a tiny bite
English
@MythThrazz I definitely dont do this with my gimbal cam.. nope. Not once.
English
i said i'd make it prove itself. here are the numbers.
prismml took qwen 3.6 27b, the model i've been calling king of the 24gb tier all month, and quantized it to 1.125 bits. every weight is a single sign bit now, the whole 27b packed into 3.5gb. it should be broken. sub 4 bit is where models turn to mush. it isn't broken. it's faster.
on my 3090, the 1 bit runs at 67.9 tok/s. the full q4 king does 40.1 tok/s on the same card. that's 1.69x, and prefill is 3.4x. the reason is clean, generation is bandwidth bound, and 1 bit weights move a quarter of the bytes per token, so less traffic means faster tokens.
then it gets stranger. i filled the context and watched the speed. at 131k tokens deep it still does 40 tok/s, which is the full model's fresh speed. the hybrid attention holds where a normal model would collapse.
and the window. the model card says 262k on device. on a single 24gb card with 4-bit kv i loaded 786k before it ran out of memory. three times their number, because the weights are so small almost the whole card is free for context. it serves at 5.2gb. a 27b model leaving 19gb free on a 24gb card.
so the compression didn't cost speed. it bought speed, depth, and a window nobody expected on consumer hardware. one thing this does not answer yet. whether it still thinks as well as the king. speed was never the real question with sub 4 bit.
the retention audit is next, same agent tasks, 1 bit against the honest q4, and we find out if 89.5% of the intelligence actually survived. that one i'm not calling early.
Sudo su@sudoingX
this lab took qwen 3.6 27b, the model i've been calling king of the 24gb tier all month, and crushed it down to 3.9gb. that's it on my screen right now, loaded now on my single 3090, using less memory than a gemma 12b. here's what they actually did, because it matters. no retraining, no new model. they took the existing weights and quantized them to binary, one sign bit per weight with a shared scale every 128, the whole 27b of it, embeddings, attention, all of it, packed into 1.125 bits. it should be broken. sub 4 bit is where models usually turn to mush. prismml claims it keeps 89.5% of the full model's intelligence. bold. i already own the real qwen 3.6 27b numbers from my own bench, so i'm going to make it prove that. same agent tasks, same tests, 1 bit against the honest q4, and we find exactly where it breaks. running it now. receipts coming.
English
@Deep_Star_Six @ZackKorman Wouldn’t it be wild if someone discovered sleepers in a Chinese model and then realized it was just a distill of a Fable sleeper?
English
