What we learned.
Plain-language guides to the measured results: what worked, what failed, and what the data can actually tell us. Each article starts with a TL;DR and links to the evidence.
syntheses
Image & video models: the quick picks
Which model to use, how long it took, quality out of 100 and one strong example from each tested configuration. No long write-up.
Open the quick reference →Images, music, and the limits of video on a 6 GB GTX 1060
153 retained files from 160 attempts: useful images, short music clips and flawed video. Learn what completed, what failed, and why a saved file is not a quality pass.
Read the finding →How fast can a personal AI setup actually be?
Small models, waiting times and a routing comparison: what the recorded laptop and CPU tests show, with the limitations made explicit.
Read the finding →dead ends
Things we tried that didn’t work. Published so the next person can skip the same dead ends — what doesn’t pencil is often more useful than what does.
Gemma on older NVIDIA GPUs: Vulkan helped; drafting did not
In the May 2026 tests, Vulkan improved generation speed by 27–47% in two configurations but slowed prompt processing. The attempted drafting paths did not deliver the hoped-for gain.
Read the dead end →More findings + dead ends shipping as we run more campaigns.