fak — Hero Benchmark Comparison (v2, SOTA-only)

Generated from tools/hero_benchmark.data.json. fak — 50-turn × 5-agent fleet serving time: 19.0 minutes (lower is better).

fak fleet serving time

19.0 min

50-turn × 5-agent, Q8

tuned SOTA cache floor

78 min

warm per-agent KV (4.1× more)

KV-cache hit rate

86.7%

vs SGLang 50–99% · =100% opt

single-stream decode, 7B

8.7 tok/s

llama.cpp 17.3 — trails, closing

Log scale (minutes) so all three bars are visible — the gap is the point. Both arms run fully live on the identical Q8 forward pass, bit-identical tokens — baseline = the tuned warm per-agent KV cache (the SGLang / vLLM floor).