Run them yourself.
A tile from the held-out test fold is already loaded and every selected model has been run on it. The graphs are the exact ONNX files the benchmark measured, downloaded and executed in your browser. Nothing leaves your machine. Add or remove architectures, or drop in an image of your own, and the table follows.
Why the browser is slower
These graphs run through WebAssembly here, which is several times slower than the native runtime the benchmark used. The millisecond column is honestly measured on your machine; it is not the benchmark's latency, and the two are labelled separately everywhere they appear.
Watch quantisation fail
Switch precision to fp32 and back on MobileNetV3-Small. Its int8 form is the one post-training quantisation destroyed, and it will cheerfully give you a confident wrong answer faster than any other model on the page.
What you are actually running
These are the benchmark's own exported graphs, not a re-implementation and not a server call. Your browser downloads each one, builds a session, runs it on the tile and releases it again. The accuracy and energy columns come from the committed measurements on an Apple M2; only the millisecond column is yours.