llama.cpp b9760 adds video input and generalized schema

llama.cpp's newest release, tagged b9760, marks a significant step toward multimodal local AI inference. The core change is a refactored input file schema that generalizes how the library accepts data. Previously limited to text and images, the server now supports an 'input_video' field, accepting raw base64 encoded video. This leverages the same modular infrastructure, making it easier to add future modalities. The commit also includes nits, fixes for Windows, and updates across all major platforms.

The implications for developers are substantial. By moving from ad-hoc input handling to a generalized schema, llama.cpp can now process video frames directly—enabling applications like real-time video analysis, multimedia summarization, and vision-language model interactions entirely on-device. The update supports various backends: CPU on Linux/Windows/macOS, GPU via CUDA 12/13, Vulkan, ROCm 7.2, OpenVINO, SYCL, HIP, and Android arm64. This ensures broad hardware compatibility while maintaining the performance llama.cpp is known for.