Key Info

A Qwen ambassador reports running Qwen3.8 Flash-Next on a single RTX 5090 with 128 GB DDR5, reaching 150 tok/s decode for programming workloads using the Strata engine and ISTA Lab's IQ3_S quantization.

Highlights

  • 150 tok/s decode on programming prompts with a single 5090
  • ~130 tok/s median on mixed prompts, about 100 tok/s on long prose
  • Setup uses 128 GB DDR5 in a non-ideal 4-stick dual-channel configuration
  • Performance attributed to Niko1221's Strata engine plus IQ3_S quant
  • Demonstrates high-throughput local inference on consumer hardware