#2165 bench: RTX 5090 + Ryzen 7 9800X3D, Qwen3.8-Flash-Next IQ3_XXS @ 131k (Strata 0.1.43.1)

说明

Community benchmark run on an RTX 5090 (32 GB, **power limit capped by the user at 450 W**, default 600 W) + Ryzen 7 9800X3D, 60.5 GB DDR5.

Suite: ./bench.sh --long on Strata 0.1.43.1 (engine 0.1.43.1, commit fb58e0db), model Qwen3.8-Flash-Next GSQ RCO IQ3_XXS, max_context 131072.

Medians: chat decode 202.1 tok/s, agent decode 223.8 tok/s, prompt 4K/16K/32K = 4473 / 6656 / 6827 tok/s, agent cache hit 68.8%.

Comparison note: the existing 2026-10-09 community folder on the same GPU+CPU pair (Swift-1.5 IQ3_XXS @ 1M YaRN, 0.1.41) reported decode 166-199 tok/s; this run adds the 131k-context IQ3_XXS data point on 0.1.43.1. README has home/user/IP scrubbed; read once before merging as usual.