llama.cpp and ggml reach version 1.0.0 with a stable API promise

The team behind llama.cpp, the open-source engine for running AI models on your own hardware, has released version 1.0.0 of the project's libraries. In a post on the Hugging Face blog, the team says both llama.cpp and ggml, the tensor library underneath it, now carry the 1.0.0 number. The team presents the release as a commitment to the developers who build on its code: from now on, changes that break compatibility will come with a major version bump.

What changes

llama.cpp lets people run large language models, including ones that handle images and other inputs, on everything from consumer laptops to enterprise servers. It relies on quantization (shrinking models so they need less memory), tuned backends for different hardware, and portable C/C++ code. The project is about 3.5 years old, and the team says many downstream applications already depend on it.

With 1.0.0, the team adopts semantic versioning rules for all ggml-org projects that use it. A major bump signals a breaking change to the public C API in the include folder. A minor bump covers backward-compatible features, backend updates, new model support or API additions. A patch bump means a bug fix with no API change. The team says it will be stricter about making API and ABI breaks visible as adoption grows. ABI refers to compatibility at the level of compiled binaries.

The two releases are libllama, with public headers in llama.cpp/include/, and libggml, with headers in ggml/include/.

Where to get it

The team lists several distribution channels. Pre-release artifacts cover Linux, macOS, Windows, Android and iOS. Docker and Debian packages serve Linux. Homebrew covers macOS and Linux, conda-forge covers Windows, macOS and Linux, MacPorts covers macOS, Winget covers Windows, and an XCFramework targets iOS and macOS. The team says it wants to expand how its software reaches developers and end users, and it points packagers to a discussion thread.

Models, hardware and testing

According to the post, llama.cpp supports more than 150 architecture variants and all mainstream speculative decoding methods, a technique that speeds up text generation. ggml runs on many kinds of hardware through dedicated backends, and each backend can be compiled as a dynamic library and loaded at runtime.

The team says it is putting more resources into its continuous integration (CI) testing. It uses Hugging Face tools, HF Buckets and HF Jobs, to scale testing for every change that reaches the master branches. It also says NVIDIA, AMD, Qualcomm, Arm, IBM and Intel provide hardware support, and it keeps adding self-hosted test runners across major platforms.

On AI-assisted coding, the team says it uses LLMs in its development work with care, and that code quality, control of the architecture and understanding of implementation details remain top priorities. The post also mentions Llama desktop apps for macOS and Windows from the same team and Hugging Face.

What to watch

Under the new rules, the major version number is now the signal to check. If you ship software on top of libllama or libggml, the team's own policy says a jump to the next major version marks a breaking change to the public C API. Minor and patch releases are meant to stay compatible.