dnveeee3

[dnveeee3 @cro33v]

Tried tencent's "1 bit" hy-3. Doesn't fit in anything I own, so it's running on everything I own. My decade-old server, my gaming PC and my laptop. 3 machines, 3 different OSes, 1 model over RPC. Spark-beating 40 tok/s at 60k context. Feels like it rivals Sonnet 4.6 in quality.

[dnveeee3 @cro33v]

The setup: 2017 Dell C4130 with 4x V100 (Linux, RPC Host), RTX 5090 (Windows, RPC worker), M4 max MacBook (MacOS, RPC worker).

Llama truly is magic, able to pipe inference not only across OSes but across Cuda to Metal and X86 to ARM.

[dnveeee3 @cro33v]

@teortaxesTex @TencentHunyuan https://t.co/comETDChbx

I am trying it now because a model performing this well under heavy quantisation is very impressive.

[dnveeee3 @cro33v]

@thsottiaux Sol has been very interesting and impressive! Between a/b testing it against Fable on some problems, it feels like a battle between strong RL vs just a bigger model resulting in sol coming up with some crazy solutions.

[dnveeee3 @cro33v]

@jun_song Using sol to configure some jank rpc inference, hope I don’t get banned

[dnveeee3 @cro33v]

@The_Guy_Space Falcon need not hover and is thus much more fuel efficient in terms of the burn needed.

However the landing legs contribute significant weight and may offset this advantage entirely.

[dnveeee3 @cro33v]

@mattshumer_ Have fun in Dario's dystopic hellscape where you'll have your door broken down for downloading GLM 5.3 off the huggingface bay or some shit while the rest of the world gets to enjoy the fruits of open source.

[dnveeee3 @cro33v]

@sama Are these per stream speeds or aggregate?

If each user gets 750 t/s that is world-shatteringly fast and honestly more exciting than the intelligence improvements brought by 5.6.

[dnveeee3 @cro33v]

@SkepticalCent @Tablesalt13 @judgesnedge My god its fucking real

https://t.co/1r7Xz8nbIK

I work in hospital ORs and they crank the AC to keep the surgeons cool under all the shit they have to wear as well as maintain positive pressure. To think there were hospitals operating outside of the 3rd world without ac... lol

[dnveeee3 @cro33v]

@iScienceLuvr It's almost poetic how they did this the same week GLM reached parity.

[dnveeee3 @cro33v]

@HallsSoHidenn @gdb They say in the announcement they believe they can grant access to the masses in the coming weeks( how benevolent) but why not delay the announcement until then??

Still the way this is all heading is very concerning :(

[dnveeee3 @cro33v]

@AnotherHowie @cursedfrog67996 @ChasersFrame 2070 super is neck and neck with the 1080ti,above is showing a base 2070.

Where I'm at the 3070 even 3070ti are about 50 dollars more than the 2070 super but offer significantly more performance. If you want maxmum value however the 3080 is it.

[dnveeee3 @cro33v]

@cursedfrog67996 @ChasersFrame 1080ti is quite a bit faster than the 2070, which can be seen outperforming the steam machine by about 10 percent here.

[dnveeee3 @cro33v]

@BenPielstick As far as I'm concerned they just killed any lead the US had internationally. If the plan is to limit these models to "US citizens" they are fucked.

[dnveeee3 @cro33v]

@RdHdSteppeson @geras256 @ducholog Yeah, because playing video games is comparable to commuting, this hypothetical is moronic.

[dnveeee3 @cro33v]

@RdHdSteppeson @geras256 @ducholog You gotta be pretty fucking lame to choose the commuter car over the f1 car if they cost the same price.

[dnveeee3 @cro33v]

@byehmuu @NormBreaker3 It's bait lol, the i5 is as default of a cpu as you can get

[dnveeee3 @cro33v]

@Blazar_Impact @O107_ @WattTheF1ux @mooreslawisdead Whatever man enjoy your steam machine, clearly you are the target audience.