Dexmal's next-generation Vision-Language-Action model (VLA) for open-world robot control.