Readme
Quants have last 8 transformer blocks (24..31) overridden to Q8_0, as per recommendation.
for the mmproj Q4/Q8 quants, Some tensors (patch embed, depthwise, down conv) are F16 (as per recommendation)
As per the comments in
https://github.com/ggml-org/llama.cpp/pull/21616 run this model with
--reasoning off
