The latest llama.cpp snapshot, build b10075, delivers a targeted enhancement for Qualcomm’s Hexagon DSP. A new CLAMP (clamping) operation has been implemented, which can improve numerical stability and performance on Snapdragon-powered devices. Users running self-hosted language models on compatible Android hardware or other Hexagon-equipped edge platforms should see efficiency gains from this low-level operator addition.
Alongside the Hexagon update, the release offers an expansive set of pre-built binaries, ensuring seamless deployment across virtually every major environment. macOS users can choose from Apple Silicon and Intel builds, while Linux options span x64, arm64, s390x, Vulkan, ROCm 7.2, OpenVINO 2026.2.1, and SYCL (both FP32 and FP16). Windows binaries cover CPU, CUDA 12 and 13, Vulkan, OpenVINO, SYCL, and HIP, with Android arm64 and iOS XCFramework rounding out the mobile front. Note that the macOS KleidiAI-accelerated build and openEuler binaries are temporarily disabled in this iteration. A dedicated UI package is also available for those preferring a graphical interface.
