The 5080 is a solid choice for this, tho you really gotta be careful with the VRAM limits. I remember when I first tried running a 32b model... I thought I had plenty of room until the context window started filling up and the whole system just chugged. It was super frustrating because I kept getting these random out-of-memory errors halfway through a generation. Make sure to keep an eye on your quantization levels because that 32b model is gonna eat up space way faster than youd expect. I would suggest sticking to a 4-bit quant if you want to keep things stable, otherwise youre definitely gonna run into some headaches once you start doing longer prompts.

> The 5080 is a solid choice for this, tho you really gotta be careful with the VRAM limits. @Reply #2 - good point! honestly tho, i'm a bit more skeptical about the 5080 for this specific build. i remember when i first started messing with llms, i bought a card based on speed rather than memory capacity. it was a huge mistake... spent more time debugging memory errors than actually using the model. from a cost-conscious perspective, i'd actually suggest looking at a used NVIDIA GeForce RTX 3090 24GB instead. for a 32b model like deepseek, that 24gb of vram is basically the sweet spot. you can run a 4-bit quantization comfortably and still have plenty of room for a decent context window. going with a 16gb card—even a fast one—is gonna force you to use much lower quants, which really hurts the model logic. i've found that having that extra breathing room makes the whole experience way more reliable. just feels safer knowing you wont hit a wall halfway through a prompt.

• 24gb is the safe minimum imo
• used 3090 saves money for more system ram
• better reliability for long prompts

I running it with my RTX 5080.