llama.cpp and ggml reach v1.0.0 with a promise on breaking changes
The open-source engine behind many local AI apps now signals breaking API changes with major version numbers. The release also drops 35 deprecated functions.
Key takeaways
- llama.cpp and its underlying ggml library both reached version 1.0.0, the ggml-org team said.
- Breaking changes to the public C API will now come with a major version bump under semantic versioning.
- The GitHub release notes say version 1.0.0 removes 35 deprecated functions and adds FP8 key-value caches and a MoE expert cache.
- The release adds support for new text, speech and vision models, and security fixes for the bundled llama-server.
- The maintainers say llama.cpp supports more than 150 model architecture variants.
llama.cpp, the open-source software widely used to run AI language models on local hardware, has released version 1.0.0 together with ggml, the tensor library it is built on, the ggml-org team said in a Hugging Face blog post. The maintainers say the main point of the release is a stable promise to developers: from now on, changes that break either library's interface will come with a new major version number.
According to the team, the project is 3.5 years old and a large number of downstream apps now rely on it, which makes "a flexible but stable API" essential. The release was published on GitHub on Nov. 13, according to the release page.
What does version 1.0 promise?
Both libraries now follow semantic versioning, according to the post. A major version bump marks a breaking change to the public C API, a minor bump covers backward-compatible features, backend updates, model support or API additions, and a patch bump is a bug fix that leaves the API alone. The maintainers say the projects "will reflect API and ABI incompatibilities with major-version bumps," referring to both the source-level interface and the binary interface that compiled apps rely on. The team adds that it will be stricter about making those boundaries visible as adoption grows.
Tbreak notes that the 1.0 label is aimed at developers and that none of the sources it reviewed says the release changes the apps people use today. The practical gain for users, it argues, is that tools built on the engine should break less often.
What changed in the code?
According to the GitHub release notes, version 1.0.0 drops 35 deprecated functions along with the rest of the deprecated public API, and moves the NextN speculative decoding API from an extension header into the main llama.h header. The notes list FP8 key-value caches, a host-memory cache for mixture-of-experts models whose CPU share can be set, and support for loading external data kept in separate GGUF shards.
On the model side, the notes list three new text architectures (Kimi-K3.1-Pro, Ripple and xing4_0), the drama-3 speech model, and image input for MageVL-Cowork and Kimi-K3.1. The bundled llama-server gets security fixes covering DNS rebinding, CORS and server-side request forgery, limits its Hugging Face token to huggingface.co hosts, cancels MCP tool calls when a client disconnects, and adds an OpenAI-compatible /v1/audio/speech endpoint. Its web interface was redesigned with named chat groups and a live-preview markdown editor.
Who maintains it, and where does it run?
The post, by Daniel Bevenius, Georgi Gerganov and merve of ggml-org, closes with thanks from "the llama.cpp team at Hugging Face" to long-time maintainers and contributors. The maintainers say llama.cpp now supports more than 150 model architecture variants and all mainstream speculative decoding methods, and that ggml backends can be compiled as separate libraries and optionally loaded at runtime for different hardware.
The post lists pre-release builds for Windows, macOS, Linux, iOS and Android, plus distribution channels such as Homebrew, Debian, Docker, Winget, conda-forge and MacPorts, and an XCFramework for Apple platforms. The team says it is expanding its test infrastructure with Hugging Face's own tools and with self-hosted test runners on hardware provided by NVIDIA, AMD, Intel, Arm, Qualcomm and IBM.
The maintainers also say they continue to integrate large language models into their development work carefully, while keeping code quality, control of the architecture and understanding of the implementation as top priorities.
Questions readers are asking
What is llama.cpp?
Open-source C/C++ software for running AI language models on a wide range of hardware, from consumer machines to servers, according to its maintainers. ggml is the tensor library underneath it.
What does version 1.0.0 change for developers?
Both libraries now use semantic versioning, so breaking changes to the public C API will only come with a major version bump. The release also removes the deprecated public API.
Does llama.cpp 1.0 change the apps I use?
Tbreak says none of the sources it reviewed says so. The release mainly gives developers a clearer signal about when an update may break their integration.
Who is behind llama.cpp?
The release post is by ggml-org maintainers and closes with thanks from the llama.cpp team at Hugging Face to long-time maintainers and community contributors.
Sources
- #llama-cpp
- #ggml
- #open-source-ai
- #local-ai
- #hugging-face
Get the day's top AI & tech stories on WhatsApp: Follow the NextFish channel
Get stories like this every morning.
Related
Open Source3 min
NVIDIA contributes OpenShell agent-security runtime to the CNCF
OpenShell joins the CNCF Sandbox and moves from NVIDIA's GitHub to a neutral organization. NVIDIA also announced Kubernetes tools for AI factories at KubeCon.