The Pulse
llama.cpp 1.0 sets versioning rules for local AI developers
The llama.cpp and ggml libraries have reached version 1.0.0, with major version changes reserved for incompatible interfaces. The release also removes deprecated functions and adds model, server and interface features.
Developers using llama.cpp and its underlying ggml library now have version 1.0.0 releases and a commitment that future incompatible interfaces will be signaled by major-version increases. The llama.cpp team at Hugging Face released both libraries on November 13, setting out versioning rules for ggml-org projects that use semantic versioning.
The project runs multimodal large language models on hardware ranging from consumer machines to enterprise servers. Its team says many downstream applications already depend on its library interfaces, making a flexible but stable API a priority.
How versions will change
The versioning rules apply across ggml-org projects that use semantic versioning. A breaking change to the public C API in the include directory calls for a major-version increase; backward-compatible features, backend updates, model support and API additions receive a minor-version increase.
Bug fixes without an API change receive a patch increase. The team also says API and ABI incompatibilities in llama.cpp and ggml will be reflected in major-version bumps.
The release itself removes deprecated public API functions, including llama_set_warmup and 35 other deprecated functions, according to the project’s release notes. The NextN API moves from llama-ext.h into llama.h, alongside a function rename.
Models and runtime changes
The team says llama.cpp currently supports more than 150 model architecture variants and all mainstream speculative decoding methods. Its underlying ggml engine uses dedicated backends, each of which can be compiled as a dynamic library and optionally loaded at runtime.
New text architecture support in the release includes Kimi-K3.1-Pro, Ripple and xing4_0. It also adds drama-3 text-to-speech, MageVL-Cowork and Kimi-K3.1 vision support, as well as Gemma 4 streaming audio input.
Other additions include OCP FP8 KV caches, a host-memory mixture-of-experts cache with a configurable CPU share, and loading of external PLE shards. On Windows, the release adds unbuffered model loading that mirrors the Linux DirectIO path.
Server and distribution
The server gains an OpenAI-compatible /v1/audio/speech endpoint and a temperature option for text-to-speech. It also cancels MCP tool calls when a client disconnects or a timeout occurs, and notifies the MCP server.
The server changes include fixes involving CORS origin handling, DNS rebinding and CORS proxy SSRF. The release adds opt-ins for unauthenticated tools and private media URLs, and restricts the Hugging Face token to huggingface.co hosts.
The web interface gains named chat groups and a live-preview Markdown editor that expands for long messages. On phones, it moves Models, MCP and Settings into full-screen pages accessible from the sidebar footer.
For distribution, the project lists pre-release artifacts for Linux, macOS, Windows, Android and iOS, alongside channels including Docker, Debian, Homebrew and conda-forge. The team is also expanding its continuous-integration infrastructure and adding self-hosted runners across hardware platforms, with hardware support from NVIDIA, AMD, Qualcomm, Arm, IBM and Intel.