Qwen3.8 125B (6B active) Flash-Next
on 4× NVIDIA V100 32GB · SGLang · 256,000 ctx
Summary
User reports prefill of about 4000 t/s and decode of about 60 t/s at 256k context on 4x V100 32GB.
Setup offloads 50GB of ngram data to system RAM.
MTP improves decode at low concurrency.