Send a correction or a benchmark

Most of what gets better here comes from someone pointing at a figure. The most useful messages:

• A benchmark run on hardware in the catalog, with the model, the quant file, the backend and a link to where it was published. Runs with a stated context depth are the easiest to place.
• A maker figure for how much memory the GPU can use, which would replace one of our labelled assumptions.
• A machine that is missing, or one still listed that its maker no longer sells.
• A model file size, context-memory figure or price that is wrong.

Prices are approximate maker list prices, so a different price in a shop today is expected. Makers: there is nothing to buy here, and the method is published in full.