What GPU is gonna give me the best performance for running DeepSeek V4 Flash locally without breaking my bank account? I'm trying to set up a local dev environment for my final project here in Seattle and I really need this thing to be fast for real-time code completions. I've got maybe $900 tops to spend on the card itself and I'm honestly torn.

I was looking at the RTX 4070 Ti Super because it has 16GB of VRAM and people say the newer architecture is better for these distilled models but then I went down a rabbit hole on some LLM discord where guys were swearing that a used 3090 is still king because of the 24GB buffer. The problem is I've never bought used hardware before and the idea of getting a card that was mined on for two years straight makes me a bit nervous lol. I've also seen some stuff about how DeepSeek models use different quantization methods that might actually favor certain tensor cores over others but the documentation is kinda thin on the V4 Flash specifically since it's so new.

I'm worried if I go with the 4070 I'll run out of memory if I try to run a long context window or if I want to have a second model running for RAG at the same time. But if I get the 3090 will the slower clock speeds actually make the "Flash" part of the model pointless? I just want something that wont lag when I'm halfway through a block of code. My semester ends in a few weeks so I need to make a decision by Friday so I can actually get it installed and working. Is the extra VRAM on the older card actually more important than the speed of the newer 40-series chips for this specific model? Or should I just save up more for something better...