What changed
On November 13, 2026, ggml-org released llama.cpp v1.0.0 and ggml v1.0.0. The maintainers now promise major-version increments for public C API and ABI incompatibilities. This release removes 35 deprecated functions, llama_set_warmup and the defrag-thold option, renames a context-state function, and promotes NextN speculative decoding into llama.h. It also adds FP8 KV caching, host-memory MoE expert caching, new model architectures and an OpenAI-compatible text-to-speech endpoint. Server changes tighten handling of unauthenticated tools, local URLs and CORS proxy requests.
Why it matters
Local AI products often embed libllama or expose llama-server to other services. A documented compatibility contract makes release planning more predictable, but the first 1.0 upgrade deliberately breaks older integrations and configuration flags. Server hardening also changes default behaviour for MCP/tool deployments and local media fetching. Builders should treat this as a tested migration, not merely an upgrade to a stable version number.
Version numbers become a compatibility contract
The maintainer announcement defines major versions for breaking changes to the public C API, minor versions for compatible additions and backend/model support, and patch versions for fixes that preserve the API. The team also says API and ABI incompatibilities will receive major-version bumps. This is a forward-looking maintainer policy, not proof that all downstream packages are already compatible.
The first 1.0 release removes old interfaces
The upstream release removes 35 deprecated llama.h functions plus llama_set_warmup and the defrag_thold context field and command-line flag. llama_set_embd_state_src becomes llama_set_ctx_state_src, and NextN speculative decoding moves from llama-ext.h to llama.h. In-process applications and scripts using these names must be rebuilt or changed; old binaries should not be assumed ABI compatible.
Security defaults can interrupt agent and proxy configurations
The release hardens llama-server against server-side request forgery, DNS rebinding, CORS proxy access to local network addresses and unauthenticated tool exposure. Non-loopback tool/MCP setups now need authentication or an explicit opt-in to unauthenticated tools. Private-address media URLs require a separate opt-in. These changes reduce accidental exposure, but can disrupt previously working trusted-LAN workflows. Avoid restoring permissive defaults without understanding the trust boundary.
New inference options are not universal speedups
The release adds OCP FP8 KV caches, a host-memory MoE expert cache with CPU-share controls, support for new model architectures and a /v1/audio/speech endpoint. FP8 cache footprint and model throughput depend on backend, context, hardware and workload. The release announcement contains capabilities, not an independently established across-the-board speed or cost gain.