Define viable.
We already have models that run on 16GB VRAM and can do thing we thought would be impossible just 5 years ago.

>>110215678
like I can do basic work or coding with it

Two weeks I believe.

>>110215684
Well you can.
It's not as good as a model 100 times the size obviously, but for anything not too challenging it's fine.

>>110215656
For me, it was roughly 2 years ago

>>110215724
we don't all have super computers dude I still use a thinkpad with upgraded ram and SSD

>>110215730
No super computer, just running MoE models of Qwen

>>110215656
Qwen3.5-4B will work fine in a 4GB card, and will work OK-ish on CPU only. There's a microsoft-tuned version (FrogNano) that is a passible Python vibecoder.

>>110215684
there are models that can write code that easily run on normal computers. they are just slower and can't do the really fancy stuff the ones attached to a supercomputer can. we might get there in two years tho.

>>110215908
we need the chinks to save us from the gigajew sam

>>110215656
Like a year ago

>>110215656
nobody listens to me when i say this, but you're never gonna have the kind of specialized compute required to do inference at any sort of acceptable or comparable speed without some serious hardware innovations. the current paradigms that underpin all this LLM bullshit were designed to take advantage of stuff that will be pro-level for decades to come. moore's law is dead, you will not have your hands on consumer-grade hardware like the hardware that's running the latest and greatest models in 5 years' time. from the perspective of the people who could change this, there's also no impetus to do so - they're happy selling $$$ worth of hardware to dumbfucks still trying to hyperscale their way out of obsolescence, why would they care about making less $ per capita by selling and catering to the consumer market, when all that'll change there is now joe schmoe down the road can generate his sick fucked up photorealistic squirrel smut or whatever

>>110215790
Yea. Any of the qwen coders run fine. Just use a low vram mode and set your instructions before you go out, or let it run while you sleep.

>>110216975
Nah you're p much wrong on every account.
1 frontier models on copy cats are having their gaps closing not increasing.
2 Quantization models are making way more stride in reducing size and costs than top of the line local models are making in increasing capabilities which was already impressive as I said.
3 People who get into ai typically go through api as a gateway but get into local. Definitely not the other way around, and they definitely won't subscribe to like 5 different services.
4 Ai is really plateuing in terms of performance leaps (not enough training data being a big one) whereas optimizations don't really have a big block to slow them down.
5 Hardware costs aren't gonna face increasing scarcity because the first we don't know the demand supply shock happened. Now tons of long term support infrastructure is being built. If anythin we're closer to the opposite happening with anticipated demand being too high compared to willing paying users.

>>110217028
quantization and other tricks work until they don't. let's say we even get down to 1 or 2 bit quantization, or some kind of mixed precision format, whatever - nevermind the fact that it'd be very difficult to truly guarantee a model would remain functional and competent at such low resolution. a model that most people would find "decent" or comparable to cloud/API offerings would still likely be several tens of billions to low hundreds of billions of weights, meaning it'd take up, best case, tens of GBs of storage. good luck. and even them, you've only fixed half the problem. even if the cost of storing the model weights was free, you have to somehow be able to load a sufficient amount of them into VRAM at once to get any sort of speed out of your inference - if it spends all its time copying from RAM -> VRAM or disk -> RAM, you're fucked.

>>110215656
We've had that for quite some time now.
Just fire up strata and you are good to go with minimal effort

>>110217868
I need it for 3d texture generation