zyguard is a GPU-accelerated local inference engine written from scratch, with a text-based chat client. It targets the LFM2.5 model family and runs models locally. It is published under the MIT licence.