Cosmos-Reason2-2B: benchmark results
Provider: Other.
Aggregate Rating 1567 ± 35, rank #1447 of 10916 rated models, from 30 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Physical AI Bench - Understanding Overall | 56.4 | Overall Score (%) | 45.8 |
| GST-Bench - Egocentric Distance (Semantic) | 7.65 | Mean relative accuracy (%; distance to a previously seen obj | 42.9 |
| GST-Bench - Top-Down Selection (Hard) | 26.4 | Accuracy (%; multiple choice of the top-down image of the ex | 38.1 |
| GST-Bench - Top-Down Selection (Easy) | 78.7 | Accuracy (%; multiple choice of the top-down image of the ex | 33.3 |
| VANTAGE-Bench - Spatial | 63.2 | Spatial (%) | 29.4 |
| VANTAGE-Bench - Temporal | 33.73 | Temporal (%) | 29.4 |
| VANTAGE-Bench - Single Object Tracking | 26.85 | Single Object Tracking (%) | 23.5 |
| GST-Bench - Top-Down Position | 17.75 | Mean accuracy (%; predicted position of the camera on the gl | 19 |
| SpaceNum | 31.6 | Accuracy (%; macro-average of 20 task cells over 3,800 multi | 17.6 |
| VANTAGE-Bench - Event Verification | 55.27 | Event Verification (%) | 16.7 |
| VANTAGE-Bench - Semantic | 59.98 | Semantic (%) | 16.7 |
| VANTAGE-Bench - Video QA | 64.69 | Video QA (%) | 16.7 |
Interactive version: theaggregate.ai/model?slug=cosmos-reason2-2b · How It Works · Data refreshed daily, snapshot 2026-11-08.