llama.cpp and ggml reach v1.0.0, bringing semantic versioning to local LLM inference

ggml-org released llama.cpp v1.0.0 and ggml v1.0.0 on November 13, 2026, according to the team's announcement on Hugging Face. The two releases give the core libraries behind local LLM inference their first stable version numbers after roughly three and a half years of rolling 0.x builds, and they signal that the public API is now treated as a supported interface rather than something that shifts from release to release.

What shipped

According to the release notes, llama.cpp v1.0.0 is the LLM inference engine itself, written in C and C++ and shipped as libllama, while ggml v1.0.0 is the machine learning tensor library underneath it, libggml. Both were tagged on the same day.

For ggml, the jump comes from build-style identifiers such as 0.9.0-dev. The project's GitHub history shows a commit introducing semantic versioning, landed together with a release-script dry run that stepped the number from 0.9.0-dev straight to 1.0.0.

How the versions will be managed

The Hugging Face post spells out the rules. A MAJOR bump is reserved for a breaking change to the public C API in the include directory. MINOR covers additions that stay backward compatible, such as a new feature, a backend update, support for an additional model family, or a new API. PATCH is for fixes that leave the API untouched. From this point on, incompatibilities in either API or ABI will show up as a major bump, which the team says makes the boundaries stricter and easier for downstream consumers to anticipate.

Distribution across eight channels

The release lists eight distribution channels, per the project: pre-release builds covering Linux, macOS, Windows, Android and iOS; Docker images for Linux; Debian packages; Homebrew for macOS and Linux; conda-forge for Windows, macOS and Linux; MacPorts; Winget for Windows; and an XCFramework for iOS and macOS. The practical effect for teams embedding llama.cpp in private tools is that the same version contract reaches them through whichever package manager they already run.

A vendor's-eye view, for now

The team positions the release as a foundation for multimodal LLM inference across hardware ranging from home machines to enterprise servers, with quantization support, optimized backends and a portable C/C++ implementation. The Hugging Face post also makes a promotional claim, that the stack is an unusually light and versatile base for cross-platform and private AI products, which should be read as the project's own assessment rather than an independent one. As the dev.to write-up notes, the release landed on the day of publication and no outside coverage had appeared yet, so the details above come from the team's Hugging Face post and the two GitHub release pages.

Why it matters

A large ecosystem of downstream applications builds on llama.cpp, spanning desktop clients, server deployments and embedded tooling. Until now, those integrations tracked an unstable 0.x line where breaking changes could arrive without a clear signal. A published versioning contract changes that calculus: teams can pin exact versions, schedule upgrades around minor releases, and treat any major bump as a review gate before it reaches production. For software this widely embedded, the predictability of the interface is itself the headline feature of v1.0.0. The code is broadly the same; what has changed is the promise about how it will evolve.