Useful speeds on an old desktop with no graphics card: what actually moves the number

An eight-year-old machine, integrated graphics, and it is genuinely usable for the sizes that suit it. What mattered, in order.

• Memory bandwidth, not cores. Generation is bandwidth-bound. Dual-channel over single-channel was the single biggest jump I measured, and it cost the price of a second stick. More cores did almost nothing.
• Mixture-of-experts models. Only a fraction of the weights are active per token, so a large model with a small active portion runs far better on a slow machine than a small dense one. This is the change that made old hardware viable at all.
• Quantisation, to a point. Four-ish bits is a large speed and memory win for a quality cost most people cannot spot. Below that it falls off quickly and stops being worth it.
• Context length, which nobody warns you about. Memory for context grows with the square of the length. Halving the window I requested was free speed and I never needed the rest.
• Threads set to physical cores, not logical. Hyperthreading made mine slower.

What did not matter: the CPU generation, the disk speed after first load, and every tuning flag I tried beyond thread count.

Expect a slow first token and a steady stream after. For chat that is fine; for anything interactive it is not.