b8880

The ggml-org team behind the massively popular llama.cpp project has released version b8880, marking a significant expansion in hardware compatibility for running large language models locally. This update addresses a key profiling bug (#22050) that resets CPU/GPU timing data when freeing context, improving performance measurement accuracy. More importantly, it delivers 28 different pre-built binaries covering everything from macOS Apple Silicon with KleidiAI acceleration to Windows with CUDA 12.4 support, Linux with ROCm 7.2 for AMD GPUs, and even specialized builds for Huawei's openEuler with Ascend AI processors.

The release represents a major step toward democratizing local AI inference across diverse hardware ecosystems. Developers can now deploy models like Meta's Llama 3 more efficiently on Windows machines with NVIDIA GPUs using CUDA 12.4 DLLs, while Linux users gain access to both Vulkan and ROCm 7.2 backends. The inclusion of Android arm64 and iOS XCFramework builds extends local AI capabilities to mobile devices, while the webgpu_context improvements lay groundwork for browser-based inference. This multi-platform approach addresses one of the biggest barriers to local AI adoption: fragmented hardware support.