Local AI was Just Made 6× Faster Running on non-AI Machine

Two days after our local-AI bearing-load experiment went live, we tuned the inference stack on the same desktop — same GPU, same model, no new hardware. Prompt processing jumped from 55 to ~630 tokens/s, generation from 8 to ~40–49 tokens/s, and the 8.5-minute bearing-internal-load-distribution run now takes about a minute, with the same verified 4,143 N answer. Here is the tuning diary, the measurement table, and the closed-form check.