The maintainers of llama.cpp, the C/C++ engine that runs language models on ordinary laptops and phones, released version 1.0.0 of its libraries on November 13, together with v1.0.0 of ggml, the tensor library underneath it. The release, announced on the Hugging Face blog by the ggml-org team, matters less for new features than for a promise: from now on, any change that breaks the public C API will come with a new major version number. For the many apps built on top of llama.cpp, that means upgrades become predictable.
What v1.0.0 changes
Until now, llama.cpp shipped as a fast-moving stream of numbered builds, and its public interface could change from one build to the next. Developers embedding the library had to watch commits closely or pin an old build.
The project now follows semantic versioning across all ggml-org projects:
- Major version: a breaking change to the public C API in the
includefolder. - Minor version: backward-compatible features, backend updates, new model support or API additions.
- Patch version: bug fixes with no API change.
The two libraries covered are libllama, the LLM inference library with its headers in llama.cpp/include/, and libggml, with headers in ggml/include/. The maintainers say both will reflect API and ABI incompatibilities, meaning changes that break source code or already-compiled programs, with major version bumps.
Why a version number matters for local AI
llama.cpp sits at the bottom of a large part of the local AI stack. Desktop apps, command-line runners and many hobby projects bundle it or a fork of it to load models in the GGUF format and run them on CPUs and GPUs. When a library at that layer changes its interface without warning, every project above it has to scramble.
The announcement says the project is now about three and a half years old and that “many downstream applications” depend on it. A stable API makes llama.cpp easier to package in Linux distributions and app stores, easier for companies to support for years and easier to audit, because a given version number now has a clear meaning.
New ways to install it
The team also listed the distribution channels it supports or is working with: pre-built release files for Linux, macOS, Windows, Android and iOS; Docker images; Debian packages; Homebrew; conda-forge; MacPorts; Winget on Windows; and an XCFramework for Apple app developers. It points to the Llama desktop apps for macOS and Windows, built by the llama.cpp team and Hugging Face, for people who do not want a terminal.
On coverage, the maintainers say llama.cpp supports more than 150 model architecture variants and all mainstream speculative decoding methods. Hardware support comes through ggml backends, each of which can be compiled as a separate dynamic library and loaded at runtime. The team says NVIDIA, AMD, Qualcomm, Arm, IBM and Intel provide hardware for self-hosted test machines, and that Hugging Face infrastructure lets it run more tests on every change.
How it compares with other local runtimes
Most people meet llama.cpp indirectly. Tools such as Ollama and LM Studio provide model downloads, a friendlier interface and their own app logic, and use llama.cpp-family engines for at least part of the actual inference. Other engines target different needs.
| Runtime | Typical use | Relationship to llama.cpp |
|---|---|---|
| llama.cpp | Embeddable C/C++ library, CLI and server for CPUs, GPUs and phones | The engine itself, now at v1.0.0 |
| Ollama | One-command local model runner with a model library | Uses ggml and GGUF models; see our Ollama guide |
| LM Studio | Desktop app for chatting with local models | Offers llama.cpp as one of its engines |
| vLLM | High-throughput serving on data-center GPUs | Separate engine aimed at servers, not laptops |
For individual users the version change is invisible. For developers it is the main news: code written against v1.x should keep compiling and running against later 1.x releases. If you run the built-in server, our guide to llama.cpp’s OpenAI-compatible server still applies, and our explainer on quantization covers why GGUF models fit on consumer hardware.
The team’s stance on AI-written code
The announcement includes a short note on how the project handles AI coding tools. The maintainers say they integrate language models into development “in a thoughtful and reasonable manner” and that keeping a high-quality codebase, staying in control of the architecture and understanding the implementation details remain their top priorities. That is a notable statement from one of the most active open-source AI repositories, at a time when maintainers across open source report a rise in machine-generated bug reports and patches.
What to watch next
- How often major versions arrive in practice, and whether downstream apps switch from pinned builds to version ranges.
- Packaging of v1.x in more Linux distributions and app stores now that version numbers carry meaning.
- Whether Ollama, LM Studio and other tools move to the stable releases.
- Support for new model architectures arriving as minor releases.
Key takeaways
- llama.cpp and ggml released v1.0.0 on November 13, 2026.
- Both now follow semantic versioning: breaking C API or ABI changes require a major version bump.
- The project supports more than 150 model architecture variants and ships through Docker, Homebrew, conda-forge, Winget and other channels.
- A stable API makes llama.cpp easier to package, support and build on for local AI apps.
- Everyday users of apps built on llama.cpp will see no direct change.
Frequently asked questions
What is llama.cpp?
llama.cpp is an open-source C/C++ library and set of tools for running large language models locally on CPUs, GPUs and phones. It uses the ggml tensor library and the GGUF model format, and supports heavy quantization so big models fit in ordinary memory.
What does llama.cpp v1.0.0 change?
The main change is a versioning promise. From v1.0.0, llama.cpp and ggml follow semantic versioning, so any breaking change to the public C API or ABI requires a new major version, while features and fixes arrive as minor and patch releases.
Is llama.cpp free?
Yes. llama.cpp and ggml are open source under the MIT license and can be used in personal and commercial projects.
