SpaceThinker-Qwen2.5VL-3B: benchmark results

Provider: Other.

Aggregate Rating 1501 ± 39, rank #2340 of 10916 rated models, from 31 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
SiT-Bench - Global Perception & Mapping20.73Accuracy (%; item-weighted over the category's subtasks)66.7
CVSBench - Viewpoint Arrow32.4Accuracy (%) on View-Arrow: choose the camera-pose arrow on 64.3
AeroGround - Relative Orientation Determination28.9Accuracy (%; Spatial Understanding: select the egocentric qu63.3
MultihopSpatial - Acc@50IoU14.4Acc@50IoU (%): a prediction counts only when the answer is r52.8
AeroGround - Region Counting34.4Accuracy (%; Spatial Understanding: count category instances43.3
MultihopSpatial - 3-Hop Ego-Centric15.9Multiple-choice accuracy (%) on the 3-hop ego-centric questi40.3
AeroGround - Reverse Sequence Reasoning17.5Accuracy (%; Spatial Reasoning: find the ground frame that d36.7
SiT-Bench - Logic & Anomaly Detection28.89Accuracy (%; item-weighted over the category's subtasks)36.7
SiT-Bench - Navigation & Planning43.11Accuracy (%; item-weighted over the category's subtasks)30
MultihopSpatial - 3-Hop Exo-Centric36.3Multiple-choice accuracy (%) on the 3-hop exo-centric questi25
AeroGround - Single-Target Matching29.4Accuracy (%; Object Alignment: select the candidate in the o20
AeroGround - Trajectory Reconstruction Selection21Accuracy (%; Spatial Reasoning: select the aerial trajectory20

Interactive version: theaggregate.ai/model?slug=spacethinker-qwen2-5vl-3b · How It Works · Data refreshed daily, snapshot 2026-11-08.