Tuning Qwen3.8-Flash-Next on 2x3090: From 17 to 41 t/s Decoding

A Reddit user detailed optimizations running Qwen3.8-Flash-Next with llama.cpp on 2x RTX 3090 plus 188GB DDR4, boosting decoding from 17 t/s to 25-29 t/s via expert cache and ultimately 41 t/s after full tuning.

2026-09-03 ~ 2026-09-04 · 2 related posts

• Qwen3.8-Flash-Next on 2x3090 + DDR4: expert cache PR lifts decode from 17 to 25-29 t/s — Extension-Bid-639 · 2026-09-03

1 near-duplicate retellings: Extension-Bid-639