VibeusQEngine 1.3.1 reached 504 tokens/s for Qwen3.8-Flash-Next in... - Vibeus
QEngine 1.3.1 reached 504 tokens/s for Qwen3.8-Flash-Next in Relative-Ant-9249’s single-stream test: 2×32 GB RTX 5090, Threadripper PRO 3975WX, 251 GB RAM, 262K context, INT4 AutoRound. Credit for QEngine goes to bpx03.
Default narrative/code decode: 370/439 tokens/s; FLASH_KV=ram: 406/504. The chart compares Strata, bucko vLLM and Infernix. Repeated prompts favour QEngine’s adaptive expert cache; earlier varied traffic measured 300–400 tokens/s. The tester’s fair comparison is roughly 1.4–1.9× Strata and 3× bucko. Prefill beats Strata at 10K but trails it at 90K.
At turn 12 of a tool-use benchmark, QEngine read the prompt in 1.75 seconds, level with Strata’s 1.7–1.9.
A 13.5 GB pinned-RAM KV pool retained all eight idle ~58K-token conversations after 400K tokens of other traffic: 1.56–2.01-second returns. VRAM retained two; six needed 8.7–9.4-second prefills. Answers remained correct. RAM decode was ~10% faster alone but ~10% slower with eight concurrent 64K streams: 124 versus 139 tokens/s, one pass.
Assistant tasks: 44/44; delegation: 6/6. The earlier rc4 coding result, 26/30, sits within Strata’s 24–26 repeat range; it was not rerun on 1.3.1. Tight VRAM, source compilation for sm_120 and limited runtime history remain costs. These are single-user measurements.
