>>108837422
just use the free claude work around.

>>108837422
Put those 2000 bucks in API credits and you'll run your shitty quantized model for several lifetimes.

>>108837440
Sauce us bro

I thought the meta was to abuse $20 subscriptions. No good anymore?

just scrape a key lmao.

>>108837422
Can you not pirate AI?

>>108837422
>invest 1500-2000 bucks
who said it’s that cheap? to get a decent model running at a decent speed you will be spending more. you’ll want a 32g card from nvidia so that’s 1k+ for shit, 3k+ for Ok, then 128G system memory for another 2k
if you want to run a small dense model the gpu is enough but the bigger moes that will be better will need the system memory too

>>108837561
I was thinking for some used z840 workstation from ebay

>>108837422
It's much worse than SOTA but also it's much better than coding fully by hand with no AI assistance.
>>108837449
Inference cost is mainly determined by the number of active parameters, so any local setup running 24/7 will almost surely come out on top. The problem is you probably aren't really going to find a useful way to run them 24/7 because those tiny models require a lot of human supervision to do anything useful.

Why yes you shouldn't pay for an AI, it's stupid totry and build a computer revolving around running one when if you have to rely on an AI to do your job, you're probably not fit to be on that job.

>>108838011
Ok mr.Snailcat

>>108838011
snailcat detected lool

>>108838128
>>108838115
I'm sorry, I'm too old and tired to care for whoever that boogeyman is.
Same shit with >>>/v/ and the Eric guy or whatever he was called.

>>108838147
> /v/ mentioned
How about you go back.

>>108838147
anon...please, stop posting...save your remaining dignity

why is the local scene full of grifters selling people lies?
>just download le qwen_3.6_distilled_opus_gpt55_mythos_ultimate_megamix.exe and run it on your 6gb 1060, it's better than le cloud models

>>108838011
It's not my job. I use it to write tools for myself and do my job faster so I can have more time for the fun stuff. Now tell me what's the point of learning how to parse gigabytes of binary data, trying to find keys and tags when AI can do that in just few prompts? Should I waste a week figuring it out? What about css? Should I take a master course because I have to draw few UI panels twice in a year?

>>108837449
>>108837422
>>108837440
If Claude is so great then why can't you use Claude to make your own free version of Claude?

>>108838208
Because I don't have enough computing power

File: file.png (61 KB, 871x562)

61 KB PNG

>>108837422
With current prices, you're better off being an APIcuck and paying a small fee for top tier models rather than paying 2-3k+ (+ electricity) for mid-tier models.

>>108838288
The electricity cost really is high

>>108837422
I've only ever used local. And you'll need to invest more than that in today's market if you want to truly locally host.

I have access to at least one if not 8x H200 cards through work which have fairly low usage.
Would a local LLM running on one of those really be on par with the latest claude/codex models?

>>108838288
>>108840240
Not sure about your area, but running something like qwen 3.6 on a local setup is still pennies a month in electricity here.

>>108838172
You should do
>>108838174
You should do