llama.cpp b9672 rolls out BoringSSL update and expanded platform builds
New release strengthens security and supports CUDA 13, ROCm 7.2, and more.
llama.cpp's b9672 release, published by GitHub Actions on June 16, marks a significant maintenance update for the open-source LLM inference engine. The headline change is the vendor update of BoringSSL to version 0.20260616.0, which brings the latest cryptographic fixes and compatibility improvements. This ensures that users running local LLMs benefit from secure TLS and hashing without external library dependencies.
Beyond the security patch, the release dramatically expands the available prebuilt binaries. AMD GPU users get ROCm 7.2 support on Linux, while NVIDIA users can choose between CUDA 12 and CUDA 13 compatibility. Intel and ARM architectures are covered on both Linux and Windows, with Vulkan, SYCL (FP32/FP16), and HIP backends included. Android arm64 and iOS XCFramework builds are also provided, enabling mobile and edge deployments. The UI assets have been refreshed, and several platform builds (e.g., openEuler) remain disabled for now. This release is a must-have for developers who need a single, secure, and portable solution for running large language models across diverse hardware stacks.
- BoringSSL updated to version 0.20260616.0 for improved security and TLS compatibility.
- Prebuilt binaries now cover macOS, Linux (x64, arm64, s390x, Vulkan, ROCm 7.2, OpenVINO, SYCL), and Windows (x64/arm64, CUDA 12/13, Vulkan, SYCL, HIP).
- Includes Android arm64 CPU, iOS XCFramework, and UI assets updates for broader device support.
Why It Matters
Enables secure, local LLM inference on more hardware—from laptops to servers to mobile devices.