Two open sub-1B models answer your prompt side-by-side, identities hidden.
The inference itself — llama.cpp compiled to WebAssembly — runs entirely in your browser.
No server, no GPU, no API keys. Your CPU is the benchmark rig.
engine llama.cpp → WASM
sampling temp 0.7 · top-p 0.9
budget $0.00
your votes stored locally
LOADING FIGHTERSdownloading…
MODEL A · ~108 MB0%
MODEL B · ~484 MB0%
Models stream from the Hugging Face CDN and are cached by your browser (OPFS) —
next visits start instantly. One Web Worker per fighter: both generate in parallel,
one thread lane each, symmetric load.