1 article tagged with this topic
Switching from F16 to Q8_0 mmproj on Gemma 4 26B adds ~30K context tokens with no vision quality loss.