b9025
New release improves ARM performance and adds Windows arm64 CUDA builds
Deep Dive
The llama.cpp project released b9025, updating KleidiAI to v1.24.0 and switching to release archives. This release provides builds for macOS Apple Silicon (with KleidiAI enabled), macOS Intel, iOS XCFramework, Linux (multiple architectures and backends like Vulkan, ROCm, OpenVINO, SYCL), Android arm64, Windows (CPU, CUDA 12 and 13, Vulkan, SYCL, HIP), and openEuler variants.
Key Points
- KleidiAI updated to v1.24.0 with ARM-optimized kernels for Apple Silicon and Android
- New Windows arm64 CPU and CUDA builds support Qualcomm and other ARM-based PCs
- Adds openEuler builds with ACL Graph for Huawei Ascend NPUs (310p/910b)
Why It Matters
Enables faster local LLM inference on diverse hardware, from Apple Silicon to Windows ARM laptops.