LFM2.5-VL-3B is Liquid AI's most capable vision-language model, with grounding, screen understanding, and function calling, fast enough to run on-device.