Urbex Qwen3-VL 4B

Model checkpoint for Urbex: Agentic Spatial Grounding for City-Scale 3D Scenes.

Urbex grounds natural-language descriptions through active spatial search in urban 3D scenes. The agent uses landmark retrieval, BEV zooming and oblique-view inspection before predicting a target box.

Paper: https://arxiv.org/abs/2610.18356

Code and inference workflow: https://github.com/AMAP-ML/Urbex

This export contains model weights and processor/tokenizer files. It does not contain optimizer state or a recorded training-step counter. Its training step cannot be inferred from the exported files.