A from-scratch, GPU-accelerated local inference engine and TUI chat client for the LFM2.5 model family.