b8838

The open-source project llama.cpp, maintained by ggml-org, has released version b8838, marking a significant expansion in its cross-platform compatibility and hardware acceleration capabilities. This update delivers 28 distinct build assets targeting everything from mobile devices to high-performance servers, with notable additions including KleidiAI acceleration specifically for Apple Silicon Macs, Vulkan API support for GPU inference on Linux and Windows, and OpenVINO integration for Intel hardware optimization. The release also formalizes support for Windows with CUDA 12.4 and 13.1 DLLs, Android ARM64 CPU builds, and specialized builds for openEuler distributions on Huawei Ascend hardware.

For developers, this means significantly broader deployment options for running Llama-family models efficiently. The KleidiAI backend promises better performance-per-watt on Apple's M-series chips, while the new Vulkan support opens GPU acceleration to a wider range of graphics cards beyond NVIDIA's CUDA ecosystem. The update also includes important refactoring like renaming 'libcommon' to 'libllama-common' for clearer library organization. With builds now available for iOS via XCFramework, Android, Windows on ARM, and multiple Linux variants, llama.cpp solidifies its position as the most portable inference engine for running large language models locally across diverse hardware environments.