llama.cpp b10278 drops Metal BF16 flag, streamlines builds
New llama.cpp release removes GGML_METAL_USE_BF16, ships prebuilt binaries for all major platforms.
ggml-org published llama.cpp b10278, a new release of the wildly popular C/C++ LLM inference engine (123k stars, 21.4k forks). The headline change is build cleanup: commit d52ec04 removes GGML_METAL_USE_BF16 from all build scripts, resolving issue #26604. This deprecation simplifies cross-compilation for Apple Silicon and Intel Macs, eliminating a flag that was causing confusion during Metal GPU builds. The release is signed with a verified GitHub GPG key, ensuring authenticity.
Alongside the code change, llama.cpp b10278 ships ready-to-use binaries for an extensive platform matrix: macOS arm64 (with and without KleidiAI), macOS Intel, iOS XCFramework, Linux on x64/arm64/s390x with CPU, Vulkan, ROCm 7.2, OpenVINO, and SYCL support, plus Windows x64/arm64 supporting CPU, CUDA 12.4/13.3, Vulkan, OpenCL Adreno, HIP, and OpenVINO. Android arm64 and openEuler builds round out the lineup. The release also points to llama.app for the official app. This update lets developers grab prebuilt artifacts and skip compiling from source, while the build flag removal reduces friction for those targeting Metal backends.
- Removed GGML_METAL_USE_BF16 from all build scripts (#26604)
- Prebuilt binaries for macOS, Linux, Windows, Android, iOS, and openEuler
- llama.cpp has 123k stars and 21.4k forks, with signed release b10278
Why It Matters
Simplifies cross-platform LLM deployment by removing a deprecated Metal flag and offering ready-to-use binaries for every major OS.