Round 16 example weights to fewer bits, then compare what you save with what you lose.
The 16 weights are an illustration, drawn in groups of 4. Size and speed use groups of 128, as Llama files do, and 16-bit embeddings. Error is the typical rounding mistake as a share of a typical weight here: a hint, not the accuracy of a model. Speed is a ceiling set by memory bandwidth, in GiB for size and tokens per second.