Gemma 4 Q8 mmproj unlocks 60K+ vision context

Quantizing multimodal projections in llama.cpp offers a “free lunch” for local inference, expanding context limits for vision tasks on constrained hardware. Shifting from F16 to Q8_0 mmproj frees significant VRAM, enabling longer context without sacrificing multimodal performance, and empirical tests suggest Q8_0 can occasionally outperform F16 in specific reasoning tasks. A fix for a related llama.cpp regression bug (post-b8660) is already approved, underscoring the rapid iteration of the open-source community.