b8947
New update prioritizes q8_0 when q4_K unavailable, improving inference reliability across hardware.
The llama.cpp b8947 release, tagged by github-actions on April 27, brings a critical improvement to model quantization handling. The core update modifies the download logic to prefer q8_0 quantization when q4_K is unavailable. This fallback mechanism prevents load failures on hardware that doesn't support the more compact q4_K format, ensuring broader compatibility across CPU, GPU, and accelerator backends. The change is particularly impactful for users running inference on diverse hardware configurations, from Apple Silicon to AMD ROCm.
The release significantly expands platform support with new prebuilt binaries for Ubuntu s390x, Android arm64 (CPU), and multiple Windows variants (CUDA 12/13, Vulkan, SYCL, HIP). Apple users get updated macOS builds for both Intel and Apple Silicon (arm64), including a KleidiAI-enabled variant. OpenEuler (310p and 910b with ACL Graph) and Linux Vulkan/ROCm builds are also refreshed. These additions make llama.cpp more accessible for production deployments on servers, edge devices, and custom hardware, reducing the need for manual compilation.
- Download logic now prefers q8_0 quantization when q4_K is unavailable, preventing load failures
- New platform builds include Ubuntu s390x, Android arm64 (CPU), and Windows CUDA 12/13/Vulkan/SYCL/HIP
- Apple Silicon builds updated with KleidiAI-enabled variant for optimized inference
Why It Matters
This release improves inference reliability across diverse hardware, making llama.cpp more robust for production AI deployments.