The curse of 64GB system RAM

A user with an R9700, RTX 5060 Ti, and 64GB DDR5 RAM, typically running Qwen3.8-27B at Q6 on the R9700, found Strata significantly improved performance for Qwen3.8-Flash-Next (QFN) IQ3_XXS. While llama.cpp achieved 21 t/s with this quant, Strata delivered around 60 t/s. Concurrent operation of QFN and Minimax H3 inference was possible with specific Strata and ComfyUI settings, maintaining 50-60 t/s for QFN despite initial NVMe read spikes from H3 weight loading.

This report uniquely details how Strata enables a 3x performance increase for Qwen3.8-Flash-Next compared to llama.cpp on the same hardware, moving it from barely usable to a daily driver.