Tools

llama.cpp reaches version 1.0, promising a stable API for local AI

The open-source engine behind much of the local AI ecosystem now follows semantic versioning, so breaking changes will come only with major releases.

The Parameter story card: llama.cpp hits 1.0 with a stable-API promise
Illustration: The Parameter

llama.cpp, the open-source project that lets people run large language models on their own hardware, released version 1.0.0 this week, about three and a half years after it began.

The milestone is less about new features than about a promise. In a post on the Hugging Face blog, the ggml-org team said both llama.cpp and ggml, the tensor library underneath it, will now signal API and ABI incompatibilities with major-version bumps. "Every new integration means more developers depending on our library interfaces and therefore having a flexible but stable API is paramount," the team wrote.

What changes

The project has adopted semantic versioning across its libraries. The rules, as published:

  • A major version bump means a breaking change to the public C API in the include directory.
  • A minor version covers backward-compatible features, backend updates, new model support or API additions.
  • A patch version is a bug fix with no API change.

Two libraries move to v1.0.0 together: libllama, for language model inference in C and C++, and libggml, the general machine learning tensor library. Developers who call functions the project had already marked as deprecated should check the release notes on GitHub before upgrading.

Where it runs

llama.cpp's appeal has always been reach. Through quantization, which shrinks models to fit in less memory, and a portable C/C++ codebase, it runs on hardware from consumer laptops to enterprise servers. The team says it currently supports more than 150 model architecture variants and the mainstream speculative decoding methods, and that hardware makers including Nvidia, AMD, Qualcomm, Arm, IBM and Intel supply machines for its testing infrastructure.

The team also listed the channels it now distributes through, including Docker, Debian, Homebrew, conda-forge, MacPorts, Winget and an XCFramework for Apple platforms, with pre-release builds for Linux, macOS, Windows, Android and iOS.

On AI-assisted development, the team said it will continue to integrate language models into its workflow "in a thoughtful and reasonable manner," while keeping control of the architecture and understanding the implementation as top priorities.

Why it matters

A large share of local and private AI, from desktop chat apps to on-device assistants, is built on llama.cpp. Until now, depending on it meant tracking a fast-moving codebase where interfaces could shift without warning. A 1.0 release with clear versioning rules makes it easier for companies to ship products on local models and keep them working, which matters as more people want AI that runs without sending their data to the cloud.

Sources

  1. llama.cpp v1.0.0: A stronger foundation for the local AI ecosystem — ggml-org on the Hugging Face blog, Nov. 13, 2026
  2. Release v1.0.0 — ggml-org/llama.cpp on GitHub
  3. dev: preparing for v1.0.0 — ggml-org/llama.cpp GitHub discussion

The Parameter reports only what its sources support. Corrections: hello@theparameter.ai.