← sources

I made a post about RAM and FOMO and now I found a good deal for seconhand RAM

source: reddit (reddit:LocalLLaMA)published 2026-10-14 10:46ingested 2026-10-15 02:42

retained text

Soooo I currently have 4\*8Gb of DDR4 3200mhz cl16 RAM modules. I usually run Swift Qwen 27B in IQ4 quant from Bartowsky (the goat himself) on my 7900XT (40tks with mtp) for daily stuff, researches and code when I'm out of work token for work related stuff.

I would love to be able to run a more efficient and more accurate model like an MoE in Q8 like Qwen Flash Next that utilizes all my hardware instead of solely my 7900XT because 20Gb of vram is just not enough for higher accuracy quants for dense models.

So I found some 2\*32Gb DDR4 3200mhz cl16 RAM modules in second-hand at 300€ and will pair it with my existing ones to get 80Gb of RAM.

\> it won't cause any issues if each type of modules is on a different channel

Do you think this is a good idea?

Are MoEs really that good compared to lower quant dense models?

What is your setup? I mean real setup that went from gaming rig to AI serving server.

Please help a man in doubts

document fetched 2026-10-15 00:46 · canonical https://reddit.com/r/LocalLLaMA/comments/1x5nzll/i_made_a_post_about_ram_and_fomo_and_now_i_found