It is basically a gutted ggml/llama.cpp, but I just had the agents build highly optimized version for two specific model quants with zero consideration or compatibility for anything else.