If you run a model on your own computer, there is a good chance llama.cpp is doing the work underneath, whether you installed it directly or through an app built on it. The project has now released version 1.0.0 of its libraries and committed to semantic versioning. That sounds like housekeeping, but for everyone who builds on it, it is a promise about when things will break.
The rules are simple and published in the announcement. A breaking change to the public C API, the headers in the include folder, bumps the major version. Backward-compatible features, backend updates, new model support and API additions bump the minor version. Bug fixes with no API change bump the patch number. The same rules apply to ggml, the machine learning engine underneath, and to the other ggml-org projects, and the team says it will be stricter about flagging incompatibilities in both the API and the ABI.
The reason is scale. In the three and a half years since it began, llama.cpp has gained a long list of dependants, and the post says it now supports more than 150 architecture variants and all mainstream speculative decoding methods. Distribution is widening too: builds are listed for Docker, Debian, Homebrew, conda-forge, MacPorts, Winget and an XCFramework for Apple platforms. Testing runs on Hugging Face infrastructure with self-hosted machines supported by NVIDIA, AMD, Qualcomm, Arm, IBM and Intel.
For the tools we review this is quietly significant. Ollama and LM Studio both draw on llama.cpp and ggml, and so do many smaller desktop apps. Until now, anyone tracking upstream had to read changelogs closely to spot an interface change. A version number that signals breakage lets those projects update more often with less risk, which should mean new model support reaches end users faster. The team at Hugging Face also ships its own Llama desktop apps for macOS and Windows.
The limits are worth stating. Semantic versioning is a commitment, not a guarantee, and it covers the C API, not every behaviour change in output quality or speed. It also says nothing about the command-line tools or server flags many people use directly. One commenter noted that processing very long prompts remains slower than people would like, and version 1.0.0 does not claim to address that.