Llama.cpp rolls out massive multi-platform update
New CUDA 13 support lands for Windows, AMD APUs get fixes, and more
The maintainers of [llama.cpp](https://llama.app), the popular open-source C++ library for running large language models (LLMs) locally, have pushed a significant update that broadens hardware compatibility and performance. Version b10472 drops fixes for AMD APUs where hipMemGetInfo now reports accurate memory instead of over-promising via MemAvailable, resolving issue #18159. The team also added CUDA 13 support for Windows (with DLLs for versions 12.4, 13.3, and preview 13.4), ROCm 7.14 for Linux, and Vulkan backends across multiple platforms.
This release spans over 20 configurations: Windows (x64/arm64 with CPU, CUDA, Vulkan, OpenVINO, SYCL, ROCm), Linux (Ubuntu x64/arm64/s390x with CPU, Vulkan, ROCm, OpenVINO, SYCL), macOS (Apple Silicon with/without KleidiAI, Intel x64), iOS XCFramework, Android (arm64), and openEuler (x86/aarch64 with 310p/910b chips). The update also includes backend improvements for SYCL FP32/FP16 and disables previously broken configurations for Intel GPUs.
- Added CUDA 13 support for Windows with DLLs for versions 12.4, 13.3, and preview 13.4
- Fixed AMD APU memory reporting via hipMemGetInfo to avoid over-promising MemAvailable
- Now supports 20+ platforms including Linux ROCm 7.14, macOS KleidiAI, and Android arm64
Why It Matters
Enables running LLMs on more hardware with better performance and accuracy for developers.