找到 1 篇关于此标签的文章
Switching from F16 to Q8_0 mmproj on Gemma 4 26B adds ~30K context tokens with no vision quality loss.