APUS Reflex vLLM response latency and historical comparisonProvided 333-request run snapshot: P50 25.58, P95 222.36, P99 275.58 milliseconds; 333 successful, zero errors. Source screenshot, raw requests not independently verified. Historical systems have different timing boundaries and cannot support speedup ratios.Fast decision responses over vLLM HTTP.9B · effort: high · vLLM HTTP API · 1× RTX PRO 6000 Blackwell 96GBP50 response25.58 msP95 response222.36 msP99 response275.58 msHTTP success333 / 333Error rate: 0% · full-response HTTP timing4B · reported P50 / ratio-derived tails: P50 18.00 / P95 100.06 / P99 154.32 msP50 · typical responseLinear scale: 0–500 ms0100200300400500APUS Reflex 4BvLLM HTTP / reported + ratio-derived18.00APUS Reflex 9B · highvLLM HTTP API / full response / snapshot25.58Jev official APITypeSafe HTTP / 240 pooled requests280.80Laya typed / full inputLocal system_one / three-run median8.68P95 · slower tailLinear scale: 0–500 ms0100200300400500APUS Reflex 4BvLLM HTTP / reported + ratio-derived100.06APUS Reflex 9B · highvLLM HTTP API / full response / snapshot222.36Jev official APITypeSafe HTTP / 240 pooled requests376.56Laya typed / full inputLocal system_one / three-run median37.66Verified HTTP reference · concurrency 1 · three rounds × 1,040 requestsMedian-of-round P50 / P95 / P99: 25.34 / 222.83 / 276.78 ms · 3,120 / 3,120 successfulDifferent timing boundaries and workloads: these bars do not establish speedup ratios.TypeSafe: pooled 240 HTTP 200 calls; mean 302.65 ms; P99 830.45 ms. Includes public-network overhead.Same host: Google ping 31.97 ms (5); TypeSafe TCP 443 181.55 ms (10); TLS 184.57 ms (10 new connections).TypeSafe ICMP: no response. API: 240 consecutive calls. TCP/TLS probes separate; no per-request subtraction.