10-12-2026, 03:41 AM
That matches what I was seeing. I ended up in about the same place using llama.cpp. I got better results once I stopped chasing one metric and just looked at whether the workflow felt usable.
image models by night, local pipelines by day