b8829
The latest update enables 2x faster inference on Apple Silicon and refactors core libraries for better modularity.
The ggml-org team behind the massively popular llama.cpp project has released a significant update with commit b8829. This release focuses on two major areas: architectural improvements and expanded hardware support. The most notable change is the renaming of the core 'libcommon' library to 'libllama-common', a structural refactor that improves code organization and modularity for developers building on the framework. This change, detailed in pull request #21936, allows the library to be built as a shared component, making it easier to integrate llama.cpp into larger applications.
Alongside the library restructuring, the update introduces KleidiAI acceleration support specifically for Apple Silicon (arm64) macOS builds. KleidiAI is a performance optimization layer that can dramatically speed up inference on Apple's M-series chips. The release also maintains comprehensive support across platforms including Windows (with CUDA 12.4, CUDA 13.1, Vulkan, and SYCL backends), various Linux distributions, and specialized builds for Huawei's openEuler OS with Ascend AI processor support. This continues llama.cpp's reputation as the most portable and hardware-agnostic LLM inference engine available.
- Major library refactor: renamed 'libcommon' to 'libllama-common' for better code organization and shared library support
- Added KleidiAI acceleration support for Apple Silicon Macs, enabling faster inference on M-series processors
- Maintains wide platform support including Windows CUDA/Vulkan, Linux ROCm/OpenVINO, and specialized openEuler builds for Ascend chips
Why It Matters
This update makes running open-source LLMs faster on Apple hardware and improves the framework for developers building production AI applications.