llama.cpp b9817 released with OpenVINO 2026.2 and platform updates

The open-source community favorite llama.cpp—a C/C++ implementation of Meta’s LLaMA model family—has dropped version b9817. This release focuses heavily on the OpenVINO backend, which is Intel’s toolkit for optimizing deep learning inference on CPUs, GPUs, and VPUs. Key changes include upgrading to OpenVINO 2026.2.1, making OpenVINO release packages self-contained to simplify deployment, and removing hardcoded compute_op_type sets for greater flexibility. Additional operator improvements enable softmax with sink input and optimize mul_mat_id conversion for large sizes, plus modifications to add_id to support 2D and 4D inputs. These changes mean faster and more reliable inference on Intel hardware.

The release also updates build configurations across major platforms. macOS users get builds for Apple Silicon (arm64) with optional KleidiAI acceleration, as well as Intel x64 and iOS XCFramework. Linux builds cover Ubuntu x64 and arm64 (CPU, Vulkan, ROCm 7.2, OpenVINO, SYCL FP32/FP16). Windows users benefit from x64 and arm64 CPU builds, as well as GPU-accelerated builds with CUDA 12.4, CUDA 13.3, Vulkan, OpenVINO, SYCL, and HIP. Android arm64 and openEuler (x86, aarch64 with ACL Graph) are also supported. The release is tagged as b9817 with a verified GPG signature, ensuring integrity. This update continues llama.cpp’s mission to bring efficient, local LLM inference to the widest possible range of hardware.