llamaperf

Qwen3.8 125B (6B active) Flash-Next

on Unknown GPU · Strata · 131,072 ctx

Tone: mixed
Oct 5, 2026
Throughput
58.0 t/s gen · 1563.4 t/s pp
Quant
IQ3_XXS (GGUF)
System RAM
256 GB

Use cases

agentic

Summary

User reports Qwen3.8 Flash-Next GSQ-RCO IQ3_XXS at up to 58 tok/s decode and 1,563.4 tok/s prefill on a 3975wx Threadripper with 256GB DDR4, CPU only. Setup is Strata with 131k context and no speculative decoding. User notes prefill could be tuned higher and MTP/DFlash should be tested, but the IQ3_XXS quant only passed 4/8 quality tests and is not safe as a local agent without supervision.