This is a PSA for people like me who tried it and hit the wall with tool calls failing left and right, so much so that harnesses like OpenCode just didn't work: There is a fix for that. You need to pass a better chat template file, which is available (I did not write it). See also this comment. To actually use it with llama.cpp, first compile llama.cpp from source, then download the chat template file I linked above, then try this (8 bit quant in this case): ./build/bin/llama-server...