• Sources: primary, discussion
  • Summary: Georgi Gerganov merged the 1.0.0 version bump on 2026-11-13. The release removes llama_set_warmup, 35 deprecated functions and defrag_thold, promotes the NextN speculative API into llama.h, adds OCP FP8 KV caches and a host-memory MoE expert cache with a configurable CPU share, and adds Kimi-K3.1-Pro, Ripple and xing4_0. llama-server gains fixes for CORS, DNS rebinding and CORS proxy SSRF, opt-in flags for unauthenticated tools and private media URLs, an OpenAI-compatible /v1/audio/speech endpoint, and an HF_TOKEN restricted to huggingface.co, with ggml also reaching 1.0.0.
  • Why it matters: Code that links llama.h must drop the removed functions before upgrading, and self-hosted llama-server deployments should pick up the new security defaults.
  • Background: llama.cpp is the C and C++ inference engine behind much local LLM serving, built on the ggml tensor library. The KV cache stores attention keys and values for past tokens, so storing it in FP8 roughly halves its memory against FP16. In mixture-of-experts models only a few expert weight blocks run per token, which is what makes caching hot experts in faster memory worthwhile.