Bench
Pick what matters, see what wins.
This ranks measured configs on the objective you choose. It does not hand you a single "best" model: the top three are labelled on the chart, every dot's rank is one hover away, and you can change objectives whenever you want.
Reading the chart. Dots are ranked by the API on the chosen objective; the labels "#1", "#2", "#3" mark the top three among the dots currently shown, and the rank in a tooltip is also among what is shown. The amber dots joined by a stepped line are the Pareto front for the two axes on screen. A hollow dot is a quantized checkpoint.
Objective and chart view. Lowest energy, highest accuracy and best all-around open on the capability plane (accuracy across, joules per correct answer up); fastest first token opens on TTFT vs accuracy. Switch the view with the buttons whenever you like; the ranking does not change.
VRAM budget is the one filter applied by the API, since it needs each config's measured load profile. The benchmark, GPU, model source, model size and tier filters apply in the browser to the ranked set. Model source and size are read from the model's name.
Chance floor. On a four-choice task a guessing model still books about 25 correct answers per 100, so its J/correct is tiny. A cell only ranks when its accuracy interval's lower bound clears the floor; unticked, sub-floor cells are drawn with a red ring.