open-source vision model

Korean team swaps bounding boxes for region selection — robot local vision models get lighter

Korean open-source team omlab releases 10B VLX-Seek 1.5: instead of outputting coordinates, the model selects candidate regions for grounding. A stead