warm latency · 2026-10-29T11:56Z · code bd17a830a2fb · google/embeddinggemma-300m-quantized · 120 held-out task prompts · `okl check --format hook --decision tiered`, a new process each
server: loaded in 3.79 s · resident memory 1221 MB
  warm    median      p95      max   (ms, end to end)
  off        120      182      255
  on         232      381      594
  added by the model, per prompt that asked it (120 of 120): median 101 ms, p95 259 ms
