Llama.cpp 1.0.0 has been released, and with it ggml 1.0.0. Version 1.0.0 means a promise: moving changes to the public C API and ABI now change the big version, and the small version grows with new models, backends, and features. Before that, the API was cleaned: 35 deprecated functions were removed from llama.h and the flag - angry child. From the new: KV cache in FP8, MoE cache in RAM, support for Kimi-K3.1-Pro and endpoint/v1/audio/speech. SSRF and DNS rebinding were closed in llama-server, and unauthenticated tools when accessing not with local adversaries are now only included in a separate flag. I wrote about 0.8.0 in October. Colleagues discussed in more detail what will break when the update is done. Channel @neuro.