llama.cpp b9964 prevents duplicate spec model downloads

The llama.cpp project, maintained by ggml-org, released version b9964 on July 11, 2024. This patch addresses a key inefficiency in speculative decoding workflows: duplicate downloads of spec models. Previously, users running multi-model inference could trigger repeated downloads of the same draft model, wasting bandwidth and time. The fix ensures each spec model is fetched only once, streamlining the setup process for local LLM pipelines. The release is signed with a verified GPG key and includes pre-built binaries for a wide range of platforms.

Beyond the core fix, b9964 rolls up UI asset updates and continues llama.cpp’s tradition of broad hardware support. Precompiled builds are available for macOS (Arm64 with optional KleidiAI, Intel x64, iOS), Linux (x64, Arm64, s390x with Vulkan, ROCm 7.2, OpenVINO, SYCL), Windows (x64 CPU/GPU, Arm64 CPU, OpenCL Adreno, CUDA 12/13, Vulkan, OpenVINO, SYCL, HIP), Android (Arm64 CPU), and even openEuler. This release is a small but meaningful step for the open-source community running language models locally, especially those experimenting with speculative decoding to speed up generation.