Best-value 24GB build: a used 3090 punches above its price

This build chases the best capability per pound in local AI, and for a lot of people that means a used RTX 3090. It is a previous-generation card, but the 24GB of memory is what matters, and second-hand it often costs less than a new 16GB card.

Who it is for

People who want to run larger models than a 16GB card allows, and who are comfortable buying used hardware to get there. It is the natural step up from a budget first build when 16GB starts to feel limiting.

The key part

The RTX 3090 brings 24GB of VRAM, enough to run models up to around 32B at a 4-bit quantisation. It is slower than current cards and draws a fair amount of power, but for holding larger models cheaply, little else competes. Pair it with a modern CPU, 32GB or more of system RAM, and a power supply with real headroom, since the card is power-hungry.

If you would rather buy new, an RTX 5080 is faster but has only 16GB, and an RTX 5090 has more memory and speed at a much higher price.

What it runs

• Qwen3 32B or QwQ 32B at Q4, with short context
• Qwen3 30B-A3B for faster generation at similar memory
• Most 7B to 14B models with plenty of room for long context

The trade-offs

It is second-hand, so condition and remaining lifespan are a gamble, and its power draw and heat are real. Two used 3090s give 48GB for those who want to go further, at the cost of a more complex, hotter build. For a single-card machine that holds large models without a flagship price, it remains hard to beat.