Honestly, grab two used NVIDIA GeForce RTX 3090 24GB GDDR6X instead of one NVIDIA GeForce RTX 4090 24GB GDDR6X. You need that 48GB VRAM total for decent local inference.

Re: Building on the earlier suggestion, be careful about the dual card setup because it really is a massive pain for airflow. Just catching up on this thread and honestly, I love your enthusiasm for local LLMs! Those 3090s are absolute beasts for the price, but you really gotta know what you are getting into with your current rig. Are you planning on upgrading your power supply and case size, or are you hoping to fit this all into your current build? That is the real kicker here.

• If you go with the NVIDIA GeForce RTX 3090 24GB GDDR6X, you get that massive VRAM buffer which is basically the gold standard for running bigger models locally without offloading to your slow system RAM.
• The NVIDIA GeForce RTX 4090 24GB GDDR6X is fantastic for speed, especially if you are doing heavy training or fine-tuning, but for just inference, the extra memory bandwidth is often wasted unless you are pushing huge batches.
• Honestly, if you have 2k to burn, you might wanna look at a high-end Corsair RM1200x Shift 1200W ATX Power Supply to keep everything stable. Getting a 4090 is great if you want to play games at ultra settings too, but if it is purely for DeepSeek, two cards usually win on capacity alone. Let me know your current power supply wattage and what case you are rocking, and I can tell you if you have a shot at fitting this hardware!

Building on the earlier suggestion, be careful about going the dual card route if you dont have the right power supply and cooling. Honestly, fitting two cards in a standard case is a total nightmare for thermals. You might want to consider the power draw before committing, since those cards run super hot. Definitely make sure your motherboard has enough PCIe lanes for that setup, or you're gonna have a bad time with speeds.