llama.cpp ships 1.0 with a stable API promise

The most boring line in this release is the important one. llama.cpp now promises not to break your code without warning.

The ggml-org team released v1.0.0 of llama.cpp and of ggml, the tensor library under it, in a Hugging Face post on Nov. 13. The post is signed by Daniel Bevenius, Georgi Gerganov and merve. The team puts the project at about three and a half years since its inception.

The version number is a contract

Both libraries now follow semantic versioning, the team says. A major version means a breaking change to the public C API in the include folder. A minor version adds backward-compatible features, backend updates, model support or API additions. A patch fixes bugs and leaves the API alone.

Both projects will flag API and ABI incompatibilities with a major version bump, according to the post, and the team says it plans to make those boundaries more visible to the projects built on top.

The release also widens how you get it. The post lists Docker and Debian packages for Linux, Homebrew, conda-forge, MacPorts, Winget, and an XCFramework for iOS and macOS, plus prebuilt artifacts for Linux, macOS, Windows, Android and iOS.

A 1.0 matters more than a new model

Yes, llama.cpp has run local models for years without a 1.0. What it lacked was a promise.

If you shipped an app on top of libllama, every update was a bet on whether your build still compiled. A versioned C API turns that bet into a policy you can plan around. Product teams need that before they put a local model in front of customers.

The reach is already wide. The team says llama.cpp supports more than 150 architecture variants and all mainstream speculative decoding methods. Each ggml backend can be built as its own library and loaded at runtime.

It says NVIDIA, AMD, Qualcomm, Arm, IBM and Intel contribute hardware support. Checks now run on every change merged to master, on CI scaled with Hugging Face Buckets and HF Jobs.

If you followed Mistral opening the weights of Large 4, tools like this decide whether an open-weight model actually runs on your own machine.

What to do with it

If you build on libllama, pin to 1.x and watch for major bumps instead of reading every commit.

If you just want local models on a laptop, the team points to Llama desktop apps for macOS and Windows. A commenter on the post asked about Linux. The thread had no answer yet.

Watch the first 2.0

A stability promise gets tested the first time keeping it costs something. Watch how long 1.x lasts, and whether the apps built on llama.cpp start pinning to it.

Boring is the feature. Build on it.