The ggml team released version 1.0 of its library on Friday, the math engine inside the popular llama.cpp. The release drops old interfaces and adds FP8, a compact number format that newer models use. Large models that wake only part of their weights now run faster on Nvidia and Apple chips.