Hugging Face releases llama.cpp 1.0 with clearer rules for breaking changes

Developers building apps on llama.cpp now have a clearer promise about how its software interfaces will change. Hugging Face’s llama.cpp team released version 1.0.0 alongside ggml 1.0.0 on November 13, 2026, committing to major-version increases when public interfaces become incompatible—a milestone aimed at the applications that depend on these libraries.

llama.cpp is software for running AI models, rather than a model itself. Its portable C/C++ implementation supports multimodal language models across hardware ranging from consumer machines to enterprise servers. The coordinated release covers two layers: libllama, the language-model inference library, and libggml, the underlying tensor library that supplies machine-learning computations.

Reading the compatibility signal

The commitment covers the public C API—the interface developers use to call the libraries—and ABI compatibility, which concerns how compiled software works with them.

That is a promise to label incompatibility, not to stop changing the interfaces.

How the version numbers will work

• Major: a breaking change to the public C API defined in the projects’ include directories.
• Minor: a backward-compatible feature, backend update, model-support change or API addition.
• Patch: a bug fix with no API change.

The team ties this policy to the project’s growth over roughly three and a half years. Many downstream applications already depend on its work, it says. Each new integration adds another developer relying on the library interfaces, making flexibility and stability joint priorities rather than separate concerns.

Moving forward, both llama.cpp and ggml will reflect API and ABI incompatibilities with major-version bumps.

The llama.cpp team at Hugging Face

One interface policy, many hardware paths

Underneath llama.cpp, ggml handles hardware differences through dedicated backends: implementations designed for different computing platforms. Each backend can be compiled as a dynamic library and optionally loaded at runtime. That design lets the underlying engine support different hardware through separate components rather than a single hardware-specific implementation.

The team says llama.cpp currently supports more than 150 model architecture variants, alongside all mainstream speculative-decoding methods. It also commits to continued support for model architectures.

Testing and packaging the foundation

The maintainers are also investing in continuous-integration infrastructure—the automated systems used to check code changes. They say Hugging Face Buckets and Jobs help scale testing capacity. With hardware support from NVIDIA, AMD, Qualcomm, Arm, IBM and Intel, they are adding more self-hosted test machines across major hardware platforms.

The development policy pairs that investment with a stated limit on how the team adopts AI coding tools. It says it is integrating language models into development thoughtfully, while keeping code quality, control of the architecture and understanding implementation details as top priorities.

For developers bringing the libraries into packaging pipelines, the release announcement lists distribution channels including Docker and Debian on Linux, Homebrew on macOS and Linux, conda-forge across Windows, macOS and Linux, and Winget on Windows. It also lists MacPorts and Apple-platform XCFramework packages, alongside pre-release artifacts covering desktop and mobile platforms.

The versioning rules apply to all ggml-org projects that use semantic versioning, not just these two libraries.