â„–41810[Quote]
File: rumir.jpg 📥︎ (32.75 KB, 339x500) cc93ec7172fc8dc88c9f3327b34d334c2c60b3ac98cd9ccc4f07cb33793864700ImgOps
>look at me, I'm OP. I spend hours coding uoptimized, trivial snca, which claude could do in 2 minutes and I love to suck COCKS. Find me on bluesky!
don't melt off the stick too fast, we want to enjoy your tantrum
â„–41909[Quote]
Use chink local ai models to set yourself free from jews and still take advantage of the tech
â„–41911[Quote]
>>41909I actually did this and just made an inbred version of Grok
â„–41915[Quote]
>>41909how do I make this work on pentium 2 pc
â„–41917[Quote]
>>41909>cede all your coding ability and freedom to a machine because you're lazy and incompetent Yeah I'm not doing that.
â„–41922[Quote]
>>41911Appalachian grok award
â„–41926[Quote]
>>41911das giga
>>41917Don't do that obviously, only use ai for time wasting tasks nophono even cares about
â„–41946[Quote]
>>41909What local models/computing power is considered good enough to not be unreliable? I have a 8gb vram gpu and 32gb ddrr4 ram but im scared any model i run will just spew out bullshit that will require hours of debugging
â„–41947[Quote]
>>41946all local models are garbage compared to cloud frontier models, even if you buy a high-end card like a 5090 for $5 gorillion dollars it won't even compare, and with that money you would be better off buying like 2 years of codex/claude $200 plan, which will get better as the tech improves
â„–41948[Quote]
>>41947I was planning on getting a claude subscription to help me trough uni and saving up for a powerful machine in the future if local tech improves anyways, thanks for the advice doe
â„–42057[Quote]
>>41947small models will get exponentially better
For like 2k you could buy some amd instincts and get about 196gb of vram
Pair that with like 512gb of ram and you are super well off
â„–42058[Quote]
>>41948LOCAL tech is improving insanely now, you only need enough vram for the active experts now kek
You can run glm 5.3 flash on a system with like 16gb of vram and 160gb+ of ram kek
â„–42059[Quote]
>>42058With good decode and prefill speeds too
(like 60t/s dec)
â„–42064[Quote]
>>42057bus link IO bottleneck doe
â„–42065[Quote]
>>42064Thats actually not a bottleneck at all really
You get almost native speeds
â„–42066[Quote]
Its essentially streaming the experts to vram (ahead of time obviously)
With 8gb of vram and 48gb of ram, I can run qwen3.8-flash-next-gsq-rco-coder (iq3_s gsq-rco (eqv to q4km, fucking crazy kek) with the non code experts stripped, maintains most of the performance)
If i had 16gb more ram, 100% could run the full iq3_S
Qwen3.8-flash-next eqv grok 4.6 according to all sorts of benchmarks and opus 4.7 / 4.8 (depends)