What this device and this browser do for the models of this site, measured in one run: the browser's features, the CPU, a model, the GPU, the storage and the line. Nothing is sent anywhere; at the end you can copy the results or open an issue with them yourself.
This browser
WebAssembly SIMD and relaxed SIMD, 64-bit memories, shared memory (software threads), WebGPU and the private file system, as the model page finds them.
CPU
The model page's own forward pass on a made-up int8 model as wide as Llama 3.2 1B (211 MB of weights a token), with 1, 2, 4 and on software threads, doubling up to the logical cores. Beside it, the CPU's ceilings: how fast it reads alone, and loops of int8 dots and of float multiply-adds.
Model
The chosen model, with and without the SIMD kernels: the table of "Measure this browser" on the model page. Without the kernels the weights take four times the memory; a browser that does not say how much memory the device has (Safari, Firefox) skips that round, and its row says so.
GPU
WebGPU: the adapter, its shaders checked against JavaScript, how fast it reads int8 weights, and a token's work of three model sizes against the CPU. Then the same token with the next token chosen on the GPU, one layer of a token as separate steps and fused into fewer dispatches, where that layer's time goes step by step, tokens generated on the GPU (the sampling too) each read back as it comes or several at once, what a token costs besides its weights, the device's ceilings (loops of multiply-adds, int8 dots or reads alone), and a prompt's tokens all at once, with the share of its ceiling each shader reaches. The weights are random bytes.
Storage
The browser's private file system: a file written 8 MB at a time, in order and far apart, then read back and checked. The file is removed at the end.
Line
A part of this site's model and a range of a model on huggingface.co: the wait for the first byte and the speed. Then huggingface.co read no faster than 1, 4 and 8 MB/s, as a slower line would deliver it.
Page memory
How much memory this page can hold before the browser refuses it or ends the tab (an iPhone ends a tab past about 2 to 3 GB, with no error). Not part of Run all: run it on its own, after you have reported or copied the other sections' results (a tab that is ended loses them). Keep this page in front until it ends: hiding it (another app, the screen locked) or leaving it stops the run. The tab may be reloaded, and the page then says how far it got.
Results
Report them as an issue(long: the issue gets a summary, and the whole results go to your clipboard: paste them into the issue)(too long even for a summary: the issue gets only a request for the results, and the whole results go to your clipboard: paste them into the issue)
This browser did not let the page write to your clipboard. The whole results are in the box below, and the page tried to select them: copy them by hand (on a phone, a long press on the text may bring up Copy) and paste them into the issue.