Skip to content

Runtime Properties

Filterable grid of per-profile runtime properties. Click a column header to sort; shift-click a second column for a secondary sort. Use the filter row to narrow profiles. Empty cells (—) indicate the runtime_props field was absent for that profile — either a managed cloud profile where capture is not applicable, or capture has not run yet for this profile.

Each row corresponds to one benchmark profile (a profile in scripts/benchmark/profiles/). The columns come from the runtime_props field on the evidence record, captured by the props-capture extractor at run time. Fields are all optional; profiles without capture data show — and sort last.

  • Extractor — which capture script populated this row.
  • Base model — the model identifier as reported by the server (not the profile alias).
  • Model quant — quantization format (e.g. int4, Q3_K_XL, 8-bit).
  • KV quant — KV-cache quantization applied by the runtime, if any.
  • Draft model / mode — speculative decoding draft model and accept strategy, when enabled.
  • Max ctx — maximum context window in tokens as reported by the server.
  • MTP — whether multi-token prediction is enabled server-side.
  • Temp — temperature from sampling_defaults (server-side default, may differ from per-request override).
2026-05-13 15:38:48 UTC · Sample data — runtime_props will be populated by the props-capture bead (fizeau-c12e6241). Fields are optional; absent fields render as — in the grid.
Profile Provider Extractor Base model Model quant KV quant Draft model Draft mode Max ctx MTP Temp