Comment by ddren

The llama models were trained with a context size of 2048. By default llama.cpp limits it to 512, but you can use -c 2048 -n 2048 to get the full context window.