llama.cpp Release b10766 Fixes Vision Input Support for DeepSeek V4
llama.cpp released build tag b10766, which updates the internal model implementation via pull request #28154 to correctly handle vision image inputs for DeepSeek V4 models. This update allows local AI users and developers to properly run DeepSeek V4 multimodal vision-language models across diverse hardware platforms supported by llama.cpp. The release provides compiled binaries for macOS, Windows, Linux, and Android across multiple acceleration backends including CUDA, Vulkan, ROCm, SYCL, and OpenVINO.
## BACKGROUND
llama.cpp is an open-source C/C++ inference framework engineered for high-performance execution of large language models on consumer hardware. DeepSeek V4 features experimental vision capabilities, allowing models to accept both image and text inputs.