Show HN: Running Gemma-4 26B at 124 tokens/SEC on a CPU, no GPU

 (apeg.dev)

10 points | by arun-prasath 9 hours ago

1 comments

  • pmb_developer 5 hours ago
    The output head byte budget is surprising. Did you try any tradeoff where the head is compressed more aggressively but experts stay mostly untouched?