llama.cpp v1.0.0 release notes

Stable API and Versioning

With the release of llama.cpp v1.0.0 and ggml v1.0.0, the project has formalized its commitment to stability for downstream developers. Moving forward, the project will adhere to semantic versioning, where major-version bumps will explicitly signal breaking changes to the public C API and ABI. This transition ensures that the growing ecosystem of applications relying on libllama and libggml can integrate with greater confidence.

Semantic Versioning Rules

  • MAJOR: Breaking changes to the public C API located in ./include.
  • MINOR: Backward-compatible features, backend updates, new model support, or API additions.
  • PATCH: Bug fixes with no API changes.

Hardware Support and Architecture

llama.cpp remains focused on providing a lightweight, portable C/C++ implementation for multi-modal LLM inference across diverse hardware. The underlying ggml tensor library is designed to support virtually any hardware through dedicated backends, which can be compiled as dynamic libraries and loaded at runtime. The project currently supports over 150 model architecture variants and all mainstream speculative decoding methods.

Distribution and CI Infrastructure

To support broader adoption, the team is expanding distribution channels for ggml-org software. Current and future distribution methods include pre-release artifacts for Linux, macOS, Windows, Android, and iOS, as well as support for Docker, Debian, Homebrew, conda-forge, MacPorts, and Winget.

Development velocity is supported by a robust CI infrastructure, utilizing Hugging Face technologies like HF Buckets and HF Jobs to ensure correctness. The project maintains a wide array of self-hosted runners across hardware platforms from NVIDIA, AMD, Qualcomm, Arm, IBM, and Intel to guarantee compatibility and performance.

Sources