13 November 2026 · 3 min read
llama.cpp 1.0 Brings Stable API and Memory Optimisations
llama.cpp and ggml reach version 1.0.0 on 13 November 2026, promising semantic versioning and breaking changes only with major-version bumps. The release removes deprecated functions, adds FP8 KV caches, and hardens the server.
What was announced
On 13 November 2026, the open-source engine llama.cpp reached version 1.0.0, alongside its underlying tensor library ggml [1][2]. The maintainers published a blog post on Hugging Face stating that from now on, any breaking change to the public C API will require a major-version bump [1]. This is a shift from the previous practice of build-number releases that gave no indication of compatibility.
The project started in March 2023 and has since become the backbone of many local AI tools [2]. Georgi Gerganov, who created llama.cpp, and his team at ggml.ai joined Hugging Face in February 2026, with the projects remaining open source under the MIT license [2][3]. At the time of the release, the GitHub repository had about 134,844 stars [2].
What's new in 1.0
The release clears out deprecated code. It removes 35 deprecated functions, the llama_set_warmup function, and the defrag_thold option [2]. The NextN speculative-decoding API moves from the experimental llama-ext.h header into the public llama.h, and llama_set_embd_state_src is renamed to llama_set_ctx_state_src [2].
Memory and speed
Several memory features arrive. A new FP8 key-value cache (options -ctk and -ctv with f8_e4m3) can halve the memory footprint of conversation context, though no speed or quality benchmarks are provided [2]. For mixture-of-experts (MoE) models, a host-memory cache keeps frequently used experts in RAM, with a configurable CPU share for missing ones. Two new parameters, moe_cache_size and moe_cache_cpu_share, control this [2].
Server and interface
The server component (llama-server) gains an OpenAI-compatible /v1/audio/speech endpoint and fixes for server-side request forgery, CORS, and DNS rebinding [1][2]. The web interface adds named chat groups and a live-preview Markdown editor. Mobile layouts now respect device safe areas and support full-screen dialogues, and the Android chat editor can open the virtual keyboard [1].
New model support
New model architectures supported include Kimi-K3.1-Pro, Ripple-65B-Preview, and xing4_0 for text, fishaudio's drama-3 for speech synthesis, and vision for MageVL-Cowork and Kimi-K3.1 [1][2].
What it means for developers
The version number is a contract. Semantic versioning means a major bump (e.g., 2.0.0) signals an API or ABI break, a minor bump (1.1.0) adds backward-compatible features, and a patch (1.0.1) fixes bugs without changing the interface [1][2]. This applies to both libllama and libggml [1].
For developers who build applications on top of llama.cpp, this removes the guesswork from updating. Previously, a build number like b12363 only indicated that the code had changed, not whether it would break existing integrations [2]. Now, a quick look at the version tells you if your code needs rework. The promise extends to the ABI as well, so compiled applications can check the version number before linking against an updated library [2]. The removal of deprecated functions means anyone compiling against the library will need to review their calls [2].
The project supports over 150 model architecture variants and all mainstream speculative-decoding methods [1][2]. ggml runs on virtually any hardware through dynamically loaded backends [1]. The team uses hardware from Nvidia, AMD, Qualcomm, Arm, IBM, and Intel for its continuous integration runners, and it integrates AI models judiciously into its own development process, prioritising control over architecture and a deep understanding of implementation details [2].
Availability
llama.cpp 1.0.0 is available now on GitHub under the MIT license [2]. Pre-release builds are provided for Linux, macOS, Windows, Android, and iOS. It can also be installed via Docker, Debian packages, Homebrew, conda-forge, MacPorts, Winget, and an XCFramework for Apple platforms [1]. An official release tag is created only when the bundled ggml matches the standalone ggml release, ensuring system-wide installations through Debian and Homebrew remain compatible [1].
The release does not change any end-user application on its own. Users of desktop apps like the Llama app for macOS and Windows, or mobile tools that embed llama.cpp, will see the benefits indirectly as developers adopt the stable API [1]. The maintainers frame the 1.0 milestone as part of a commitment to AI sovereignty and privacy, keeping the technology from being controlled by a small number of companies [1].
Sources
- llama.cpp 1.0 reaches stable release for local AI
- llama.cpp 1.0 arrives with a stable API and 35 fewer deprecated functions
- llama.cpp reaches version 1.0 and promises a stable API for local AI apps
- llama.cpp
- local ai
- api stability
- ggml
- open source
Researched and drafted with AI from the sources above.
Learn to build with AI
Hands-on tutorials on Cursor, Claude Code, AI agents and more.