Cosmos-Reason2-8B: benchmark results
Provider: Other. Access: Open.
Aggregate Rating 1630 ± 20, rank #754 of 10916 rated models, from 121 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| RepProbe Real - Exact | 17.5 | Exact (%) | 100 |
| RepProbe Real - MAE | 2.89 | MAE (lower is better) | 100 |
| RepProbe Real - Normalized MAE | 0.43 | Normalized MAE (lower is better) | 100 |
| RepProbe Real - OBO | 47.5 | OBO (%) | 100 |
| RepProbe Real - OBO Many | 22.2 | OBO Many (%) | 100 |
| RepProbe Real - OBO Mid | 43.1 | OBO Mid (%) | 100 |
| TAR-Bench - Multiple-Choice QA with Explanation | 76.2 | Scaled BERTScore F1 (%; a multiple-choice answer with an exp | 100 |
| Physical AI Bench - Understanding Overall | 65.7 | Overall Score (%) | 91.7 |
| SiT-Bench - Logic & Anomaly Detection | 40.22 | Accuracy (%; item-weighted over the category's subtasks) | 88.3 |
| EgoDyn-Bench | 39.9 | Balanced accuracy (%) of semantic ego-motion questions (14 t | 86.7 |
| MemLens (128K) - Temporal Reasoning | 43.81 | Accuracy (%) on the temporal reasoning (TR, 194 questions) a | 84.8 |
| RepProbe Real - OBO Few | 62.1 | OBO Few (%) | 83.3 |
Interactive version: theaggregate.ai/model?slug=cosmos-reason2-8b · How It Works · Data refreshed daily, snapshot 2026-11-08.