AnalysisDevelopersOctober 27, 2026

LocalLLaMA users benchmark inference engines for Qwen 3.8 Flash Next

Read original source →github.com

One user reports Strata lifting Qwen 3.8 Flash Next decode from 70-80 tok/s on llamaserver to 150 tok/s. A separate test hit 504 tok/s decode on 2x RTX 5090 with QEngine 1.3.1, and a third engine, sfnx, claims ~400 tok/s on 3x V100 32GB.

How this story unfolded

11 days · 0 reports · 5 community posts · from Oct 23

  1. Oct 23
  2. Oct 27
  3. Oct 28
  4. Oct 31
  5. Nov 3

More stories today

Open the live feed