llama.cpp v1.0.0 and ggml v1.0.0 remove deprecated APIs while adding FP8 key-value caches, a host-memory Mixture of Experts cache, new text and multimodal models, hardened server security, and a redesigned web interface.