llama.cpp b9959 adds CUDA 13, ROCm 7.2, and broad platform support
New release supports CUDA 13, ROCm 7.2, and ARM/S390X Linux builds.
The latest version of llama.cpp, b9959, extends its already impressive hardware compatibility. Key additions include support for CUDA 12 and CUDA 13 (with separate DLL builds), ROCm 7.2, and SYCL FP16/FP32. The update also introduces Linux builds for s390x (IBM Z) and ARM64 with Vulkan, plus openEuler builds with ACL Graph optimizations for Huawei Ascend 910B/310P. Windows gets new OpenCL Adreno and HIP backends, while macOS Apple Silicon and iOS XCFramework remain supported.
This release underscores the project's commitment to democratizing local LLM inference. By supporting older and newer hardware stacks—from gaming GPUs to enterprise ARM servers—llama.cpp b9959 enables developers and hobbyists to run models like LLaMA, Mistral, and Gemma on nearly any device. The addition of UI assets further simplifies deployment for self-hosted AI assistants. As local AI gains traction, such broad platform support reduces dependency on cloud APIs and speeds up on-device experimentation.
- Adds CUDA 12 and 13 DLL builds, ROCm 7.2, and SYCL FP16/FP32 for AMD/Intel GPUs.
- New Linux builds for s390x, ARM64 Vulkan, and openEuler with ACL Graph (Ascend 910B/310P).
- Windows now includes OpenCL Adreno (Qualcomm) and HIP (AMD) backends for broader GPU support.
Why It Matters
Enables running local LLMs on diverse hardware—from gaming GPUs to ARM servers—reducing cloud dependency.