Definition
Semantic versioning is Semantic versioning is a numbering scheme, MAJOR.MINOR.PATCH, where each number tells you whether an update breaks existing code, adds features safely, or only fixes bugs.
Software updates arrive with numbers like 1.4.2. Semantic versioning, often shortened to semver, is a rule for what those three numbers promise. It is the reason a developer can read a version bump and know, before installing anything, how risky the upgrade is.
A semver version has the form MAJOR.MINOR.PATCH. The official specification describes each part this way:
Think of it like a restaurant menu. A patch fixes a typo. A minor adds a dish. A major swaps the whole menu, so your usual order may be gone.
The specification says major version zero (0.y.z) is for initial development, and that anything may change at any time. Version 1.0.0 defines the public API. So a project reaching 1.0 is making a promise: from here on, breaking changes will be announced with a new major number.
Much of the AI stack is open source software that other software depends on: inference engines, agent frameworks, SDKs, model-serving tools. If your app depends on one of them, semver is what lets you set a rule such as "accept any 1.x update" and trust that it will not break your build.
Without that promise, every update is a gamble. Teams end up pinning an exact version and never upgrading, which means they also skip security fixes.
Semver is a promise by the maintainers, not something enforced by the computer. A project can break its own rules by accident, and what counts as the "public API" has to be defined by the project. For libraries with a C interface, that also includes the binary compatibility of the compiled code, which is easy to break without noticing.
Semver also says nothing about how good a release is. A model, a dataset or a prompt library usually does not follow it, because "backward compatible" is hard to define when behavior changes statistically.
Versioning is a cousin of model deprecation, where a provider retires a model on a schedule you have to plan around. It also matters for model context protocol servers and other tools that agents call, since a breaking change in a tool can quietly break an AI agent that depends on it.
The practical habit is simple. Read the major number first. If it did not change, the upgrade is meant to be safe. If it did, read the release notes before you update.
MAJOR changes when a release breaks compatibility, MINOR when it adds features without breaking anything, and PATCH when it only fixes bugs, per the semantic versioning specification.
It marks the point where the public API is defined. Versions 0.y.z are for initial development, when anything may change at any time.
Anthropic and Theorem say Claude wrote machine-checked proofs that 239 fuzz entry points across 123 open-source libraries never crash on inputs under 64 MB, finding hundreds of bugs on the way. AI is finally doing the slow, boring half of security.
The ggml-org team released llama.cpp v1.0.0 on Nov. 13 with semantic versioning for its C API, so breaking changes now come with a major version bump. For anyone shipping local AI, that promise matters more than any new feature.