AnalysisDevelopersOctober 27, 2026

Strata and forks push Qwen3.8-Flash-Next to 504 tok/s decode

Read original source →github.com

QEngine 1.3.1 hit 504 tok/s single-stream decode on 2x RTX 5090, ~1.5-1.9x faster than the next-best engine tested. Strata forks Basalt and tandem.cpp report 275 tok/s peak and 97 tok/s on a single 4090 respectively.

How this story unfolded

11 days · 0 reports · 8 community posts · from Oct 23

  1. Oct 23
  2. Oct 25
  3. Oct 27
  4. Oct 28
  5. Oct 29
  6. Oct 31
  7. Nov 2
  8. Nov 3

More stories today

Open the live feed