b8429

The Llama.cpp project, a leading C++ implementation for running Large Language Models (LLMs) efficiently, has released a significant new commit (b8429). The core technical addition is a new function, `clip_graph::build_mm()`, which integrates CLIP (Contrastive Language-Image Pre-training) capabilities. This marks a pivotal step for the framework, moving it from a text-only inference engine to a multimodal one. Now, models running through Llama.cpp can process and understand both images and text in tandem, enabling applications like visual question answering or generating descriptions from pictures.

Alongside this major feature, the release significantly broadens hardware compatibility. The project now provides official pre-built binaries for a wider array of systems, including macOS on both Apple Silicon (arm64) and Intel (x64) architectures, multiple Linux distributions with support for CPU, Vulkan, and ROCm backends, and expanded Windows options. Notably, Windows users gain access to binaries for CUDA 12.4 and the newer CUDA 13.1, as well as experimental builds for Vulkan, SYCL, and HIP. This expansion lowers the barrier to entry, allowing developers and researchers to deploy multimodal AI on their existing hardware without complex compilation steps.