OM-Lab's VLX-Seek-1.5-10B nails visual grounding for robots and drones

VLX-Seek-1.5-10B is an open-source 10B model built for fine-grained perception and visual grounding in embodied scenarios like drones, robots, robotic dogs, surveillance cameras, and inspection systems. Instead of generating bounding-box coordinates, it treats candidate regions as addressable entities and performs localization through region retrieval and reference—letting the model select, compare, and reason like a language model. It features stronger visual perception, faster inference via OPN proposals and more Linear Attention layers, and explicit absent-target rejection with hard-negative training and a None output format to reduce hallucinated grounding. The VLX-Seek 1.5 family is planned in 0.6B, 3B, and 10B sizes, with this card covering the 10B checkpoint. According to the article, the model still relies on candidate regions and needs its post-processing pipeline to map region tokens back to image coordinates.