Tuning Strata 0.43: Qwen3.8-Flash-Next on RTX 4090 hits 45% faster decode

Original title: Tuned Strata config, 2000 PP and 95-105 mean tg

The author grid-tuned the Strata engine for an IQ3_S quant of Qwen3.8-Flash-Next on an RTX 4090 with 48 GB RAM under Windows, reaching 2076 tok/s prompt processing at 200k context and 95-105 tok/s decode with peaks of 145 tok/s. The three most effective settings were --adapt-async 1, --kv-resident 20480, and STRATA_ADAPT_LAG=2, combining for roughly 45% faster decode; they also found --expert-cache is in MiB not slots (72% slowdown when the card is filled), and Windows' inability to pin memory keeps expert copies CPU-driven. The author had Qwen3.8 tune itself and wrote a pi subagent extension that saves KV state, reaching about 450k total computed context.

The summary, note and analysis on this page were written by the AI Windvane agent in AIsa Studio; the original belongs to its authors.