A library update can add support for a model while also forcing changes in the app that runs it. For teams building local AI into a product, knowing when an upgrade will require integration work makes a difference. On November 13, the llama.cpp and ggml maintainers released version 1.0.0 of both libraries and set out how they will flag breaking changes to their public interfaces.

Llama.cpp provides libllama, the inference library used to run language models, including multimodal models, on hardware ranging from personal computers to servers. The ggml project provides libggml, which performs tensor operations used during inference. The projects publish separate releases for llama.cpp and ggml, and both now follow semantic versioning rules.

What the version number tells integrators

Breaking changes to the public interface trigger a major version bump. Backward-compatible additions, including model support or API extensions, go into minor releases. Bug fixes that leave the API intact go into patch releases. The maintainers also commit to signaling binary incompatibilities with a major bump, which matters to compiled applications that load these libraries.

The difference becomes tangible when a team ships the same application on several platforms. An API signature change can require source changes. A binary incompatibility can mean rebuilding or redistributing the app. The version number gives teams a signal to budget for that work. The policy is anchored in public interfaces declared in the include/ headers. Integrations that rely on internals need to assess those dependencies separately.

What still belongs in the test suite

Versioning helps teams plan integration work. Each application, however, still has its own mix of models, quantization settings and hardware. Teams packaging the library can check whether inference outputs remain acceptable and performance stays within their targets on the devices they support.

For a team pinning a commit because upgrades are hard to predict, 1.0.0 offers a starting point for moving to numbered releases. Trace which calls rely on the public API and try an end-to-end upgrade. That will show where the release policy can guide maintenance and where the product still needs its own checks.

Sources. Maintainers’ announcement, llama.cpp release and ggml release.