
Gemma 4 WebGPU Kernels 255 tok/s by x/@xenovacom
We need more of this, 100+ T/s on dense models is the difference between defaulting to Claude/Codex for everything vs having a local private model doing most of the heavy lifting and only reaching for frontier for heavy intelligence work.
https://x.com/xenovacom/status/2065656427117437213