The llama.cpp and ggml open-source projects have released version 1.0.0, a milestone for the libraries that many local AI tools now build on. The ggml-org team, which includes Hugging Face staff, says the release follows 3.5 years of development and reflects a growing number of downstream applications that depend on the code. llama.cpp provides LLM inference in C and C++, while ggml is the tensor library underneath it. Both publish public headers that developers compile against, so the team says it is committing to clearer version boundaries.
Under new semantic versioning rules, any breaking change to the public C API will trigger a major version bump. Backward-compatible features, backend updates and new model support will ship as minor releases, and bug fixes with no API change will ship as patches. The team says it will make these API and ABI boundaries more visible to downstream users as adoption grows. Distribution is also expanding. Pre-release artifacts are available for Linux, macOS, Windows, Android and iOS, and packages are listed through Docker, Debian, Homebrew, conda-forge, MacPorts, Winget and an XCFramework for iOS and macOS. The project says it supports more than 150 model architecture variants and the mainstream speculative decoding methods. Its ggml backends can be compiled as dynamic libraries and loaded at runtime across a wide range of hardware.
The team also described its development process. It said it uses LLMs in its own development cycle, while prioritising code quality and a clear understanding of the implementation. It has expanded continuous integration using Hugging Face Buckets and Hugging Face Jobs, and it is adding self-hosted runners with hardware support from NVIDIA, AMD, Qualcomm, Arm, IBM and Intel. For Australian businesses and professionals, the release changes little immediately, since it is infrastructure for developers rather than a consumer product. Running models locally can keep prompts and documents on hardware an organisation controls, which may simplify handling of personal information under the Privacy Act 1988, but it does not remove obligations under the Australian Privacy Principles. Most non-technical teams will see no direct effect. Engineering teams evaluating private, on-device inference may find it worth testing, pinning versions and reading release notes before upgrading.
Key Points
llama.cpp and ggml reach v1.0.0, with future major-version bumps signalling API and ABI breaks
Semantic versioning rules: MAJOR for breaking C API changes, MINOR for compatible features and model support, PATCH for fixes
Packages now listed across Docker, Debian, Homebrew, conda-forge, MacPorts, Winget and XCFramework
Project supports more than 150 model architecture variants and mainstream speculative decoding methods
CI scaled on Hugging Face Buckets and Jobs, with self-hosted runners backed by major chip vendors