VLX-Seek 1.5 10B Released

VLX-Seek 1.5-10B is a 10B-scale open-source vision-language model targeting edge environments such as drones, robots, surveillance cameras, and inspection systems.

Instead of directly generating coordinate numbers, it converts candidate regions into language-referenceable entities, performing visual localization via a region retrieval and reference approach where the model selects, compares, and points to regions.

Key features include:

• Embodied visual grounding optimized for real-world scenes from drone, surveillance, and robot perspectives
• A powerful auxiliary vision tower and improved vision-language alignment
• Improved inference speed and memory efficiency through OPN proposal generation and Linear Attention
• Suppression of false positives for absent objects via hard-negative rejection training and None output
• A multi-scale model family consisting of 0.6B, 3B, and 10B variants

Technical details and execution code are available in the om-ai-lab/VLX-Seek GitHub repository.