A budget build using a Tesla M40 24 GB GPU and DDR3 memory achieves ~20 tokens/s inference speed with the Qwen 3.8 Flash Next IQ3 model. The post provides rough benchmarks for various flash models and notes that the M40, a 2015 Kepler‑based GPU, can be sourced for about $50 on eBay. It serves as a reference for constructing a cheap DDR3‑based server PC for local LLM experimentation.