so i’ve been really diving into the whole deepseek ecosystem lately and with v4 flash coming out i really want to move away from using their api and just host it myself. i do a lot of freelance dev work from my apartment in seattle and i’m tired of worrying about privacy or rate limits when i’m trying to churn through code. i’ve got about $1800 maybe $2k if i push it saved up for a gpu upgrade specifically for this build. my office is basically a converted closet so i really can’t have a machine that puts out enough heat to melt my face off or sounds like a jet engine taking off every time i run a prompt.

im currently looking at three different paths and i’m kinda stuck. first option is just grabbing a used rtx 3090 off ebay for like 700 bucks. the 24gb of vram seems like the gold standard for these models but the power draw scares me a bit since my psu is only 750w and like i said the heat is a real issue in this small space. then i was thinking about maybe doing two rtx 4060 ti cards - the 16gb versions. that would give me 32gb of vram which sounds insane for the price but i’ve heard mixed things about how well deepseek handles multi-gpu setups on consumer hardware without nvlink. is the latency gonna kill me?

the third way is just biting the bullet and getting a single 4080 super. it’s way more efficient and newer but only has 16gb vram. im worried v4 flash might need more than that to run at a decent speed without heavy quantization which i’d like to avoid if possible. i need to get this ordered in the next two weeks so i can have it up and running for a big project starting next month. im really trying to find that sweet spot where i dont have to compromise too much on the context window either because some of these repos i work on are huge.

so what do you guys think is the best way to go for the best balance of speed and power efficiency here? is the 24gb on the older 3090 still king or should i go with the newer dual card setup...