Budget first build: 16GB for under the price of a flagship GPU

This build is for someone trying local AI for the first time who wants a real, usable setup without spending flagship money. The goal is enough memory to run genuinely useful models, at the lowest sensible cost.

Who it is for

People who are curious, want to run models on their own hardware, and would rather prove the idea works for them before investing more. It is also a fine long-term setup for anyone whose needs sit in the 7B to 14B range.

The key part

Everything here rests on the GPU, and the choice is the RTX 4060 Ti 16GB. It is not the fastest card, its memory bandwidth is modest, but the 16GB is what counts: it runs models that simply will not fit on cheaper 8GB and 12GB cards. If your budget is tighter still, a used RTX 3060 12GB is a cheaper way in, at the cost of some headroom.

The rest is an ordinary desktop: a modern mid-range CPU, 32GB of system RAM, and a power supply with enough headroom for the card. None of it needs to be exotic.

What it runs

Comfortably, at good quantisations:

• Qwen3 8B or Qwen3 14B for general use
• Qwen2.5-Coder 7B for coding
• Smaller image and speech models

Use the hardware matrix with 16GB entered to see the full list.

The trade-offs

Generation is slower than a high-end card, and 16GB rules out the larger models except at low quantisations. That is the honest cost of the low price. For a first setup, or for anyone whose work fits comfortably in this range, it is a genuinely capable machine and the least risky way to start.