b9090
Local LLM powerhouse llama.cpp b9090 brings BoringSSL security fix and GPU backend support.
The ggml-org/llama.cpp project, the go-to library for running large language models locally on consumer hardware, has tagged its latest release: b9090. This release focuses on infrastructure updates, most notably updating BoringSSL to version 0.20260508.0 (commit #22839), which enhances cryptographic security for network-dependent operations like token streaming or model downloads. While no new model architectures or quantization methods were added, the release solidifies llama.cpp's compatibility across a wide range of platforms.
The highlight is the extensive build matrix: macOS users get Apple Silicon (both regular and KleidiAI-enabled) and Intel builds; Linux covers CPU, arm64, s390x, plus GPU backends like Vulkan, ROCm 7.2, OpenVINO, SYCL (FP32/FP16); Windows adds CPU, arm64, CUDA 12 & 13, Vulkan, SYCL, and HIP; Android gets arm64 CPU; and openEuler offers x86 and aarch64 with Ascend ACL Graph support. This ensures developers can choose the exact backend matching their hardware, from cloud VMs to edge devices, all while benefiting from the latest security patches.
- BoringSSL updated to 0.20260508.0 for improved network security
- Builds available for macOS, Linux, Android, Windows, and openEuler with multiple GPU backends
- Supports CUDA 12 & 13, Vulkan, ROCm 7.2, OpenVINO, SYCL, HIP, and Ascend ACL Graph
Why It Matters
Developers running LLMs locally gain a secure, cross-platform foundation with diverse hardware acceleration options.