Guides
Short explainers behind the numbers in the benchmark table and the advisor: what actually limits decode speed, what a measurement needs to be trustworthy, and what a quantisation label costs.
- How to measure tok/s in a way someone else can use
What to record alongside the number so that your measurement can be verified, compared, and accepted here. - Quantization, in the terms this site uses
What the bits-per-weight number means for memory, speed and the comparability of two benchmark rows. - Why memory bandwidth predicts decode speed
Token generation reads the active weights from memory once per token, which makes bandwidth the ceiling and compute mostly irrelevant.