llama.cpp b9992 refactors the CUDA MMQ (Multi-Matrix Quantization) kernel configuration and fixes the Blackwell config — the GPU architecture behind NVIDIA’s RTX 5090 and 5080.
Key changes in this build:
- CUDA MMQ refactor (#24127) — removes legacy code, fixes Blackwell kernel config
- CUDA 13.3 DLLs now ship for Windows x64 alongside existing CUDA 12.4 builds
- ROCm 7.2 support added for Ubuntu x64
- SYCL FP16 path available for Ubuntu x64
- OpenCL Adreno support for Windows arm64
- ACL Graph mode for openEuler Ascend 910b
The MMQ kernel refactor is the substantive change. MMQ kernels handle quantized matrix multiplication on NVIDIA GPUs — the hot path for inference with GGUF models. Removing legacy config paths and fixing Blackwell means RTX 50-series owners get correct kernel selection instead of falling back to slower generic paths.
NOTE
Two platforms are explicitly DISABLED in this build: macOS Intel (x64) and openEuler (generic). If you’re running llama.cpp on an Intel Mac, no prebuilt binaries ship for you this release.
The build matrix covers 30+ target configurations across macOS, Linux, Android, Windows, and openEuler — including Vulkan builds for Ubuntu and Windows, and CPU-only builds for Ubuntu s390x.
Grab prebuilt binaries on the b9992 release page. Windows CUDA 13 users should pull the CUDA 13.3 DLL variant; everyone else on CUDA 12 stays on the 12.4 build.