llama.cpp's b10280 release patches subprocess.h, expands platform support

llama.cpp, the prolific open-source C++ library by ggml-org for running large language models locally, has dropped a new release tagged b10280. With 123k stars and 21.4k forks, it's one of the most relied-upon tools for on-device AI inference. This update focuses on vendor patches for subprocess.h (issue #26606), which addresses process-handling compatibility issues that affected certain builds, particularly on Windows. The release was published on August 5, 2025, and signed with a verified GPG key, underscoring its official status.

The release ships a dizzying array of precompiled artifacts across operating systems and acceleration backends. On macOS, users can grab Apple Silicon builds, including a KleidiAI-enabled variant, plus iOS XCFramework. Linux coverage includes Ubuntu x64/arm64/s390x with CPU, Vulkan, ROCm 7.2, OpenVINO, and SYCL (FP32/FP16) options. Windows users get x64/arm64 CPU builds, plus CUDA 12 (12.4 DLLs) and CUDA 13 (13.3 DLLs), Vulkan, OpenVINO, SYCL, and HIP. Android arm64, OpenCL Adreno, and openEuler with Ascend ACL Graph support round out the lineup. UI assets are also included. This effectively means developers can run llama.cpp on nearly any hardware configuration, from cost-constrained edge devices to high-end GPU clusters.