THE ESSENTIALS IN 30 SECONDS

  • llama.cpp (libllama) and ggml (libggml) both reached v1.0.0 on November 13, 2026.
  • Breaking changes to the public C API or ABI will now require a major version bump under semantic versioning.
  • The release removes the deprecated public API, including 35 deprecated functions, and adds FP8 KV caches.

llama.cpp, the C/C++ engine widely used to run language models on laptops, phones and servers without a cloud connection, has reached version 1.0.0. The ggml-org team tagged the release on November 13, 2026, together with v1.0.0 of ggml, the tensor library underneath it. The announcement was published on the Hugging Face blog, with Georgi Gerganov, who started both projects, among the listed authors.

A versioning promise rather than a feature drop

The main change is a compatibility policy. Both libraries now follow semantic versioning across ggml-org projects: a major version increase marks a breaking change to the public C API, a minor version covers backward-compatible features, backend updates, new model support or API additions, and a patch version is reserved for bug fixes that leave the API untouched. The policy applies to the public headers in each project's include directory, and the team says incompatibilities in either the API or the binary interface (ABI) will only arrive with a new major number.

For developers who embed libllama in desktop apps, mobile apps or language bindings, this matters more than any single feature. Until now the project moved quickly through numbered builds, and downstream maintainers often had to track changes to function signatures by hand. Coverage of the release describes the tag as a signal for downstream teams rather than a freeze on development.

What else ships in 1.0

To start the stable series cleanly, the release removes the deprecated public API, including 35 deprecated functions. Reported additions include FP8 key-value caches, a host-memory cache for mixture-of-experts models, an OpenAI-compatible speech endpoint in the bundled server, and fixes for server-side request forgery, CORS and DNS rebinding issues.

The team also points to the project's breadth: support for more than 150 architecture variants, mainstream speculative decoding methods, backends that can be loaded as dynamic libraries at runtime, and packages for Docker, Debian, Homebrew, conda-forge, Winget and other channels. Continuous integration runs partly on self-hosted machines provided by hardware vendors including NVIDIA, AMD, Qualcomm, Arm, IBM and Intel.

Context

llama.cpp began in March 2023 and is distributed under the MIT licence. The value of the new policy will depend on how consistently it is applied over the coming releases, and on whether the many bindings and apps built on libllama start pinning to major versions.