Cosmos3-Edge places in the top three on PAI-Bench.
NVIDIA Jul 2026 Open weights
Its best score on each benchmark, and where that ranks among every model with a score there. We rank each model once, at its best setting.
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| IFBench | 20.8% | #79 of 80 | Lab model card | |
| HallusionBench | 40.7% | #31 of 31 | Lab model card |
| Benchmark | Score | Rank | Source | |
|---|---|---|---|---|
| CVBench | 84.9% | #7 of 15 | Lab model card | |
| RBench | 55.1% | #6 of 12 | Official leaderboard | |
| PAI-Bench | 81.1% | #3 of 6 | Official leaderboard | |
| VSI-Bench | 59.2% | #4 of 7 | Lab model card | |
| CountBench | 89.9% | #19 of 28 | Lab model card | |
| RealWorldQA | 73.3% | #26 of 38 | Lab model card | |
| MVPBench | 53.3% | #3 of 4 | Lab model card | |
| Physics-IQ | 34.1% | #3 of 4 | Official leaderboard | |
| Where2Place | 55% | #3 of 4 | Lab model card | |
| RefSpatial-Bench | 48.4% | #14 of 16 | Lab model card | |
| Physics-IQ Verified | -7.51 | #9 of 10 | Official leaderboard | |
| MMBench | 76.6% | #10 of 11 | Lab model card | |
| DocVQA | 86.8% | #33 of 36 | Lab model card | |
| InfographicVQA | 60.1% | #24 of 26 | Lab model card | |
| RefCOCO / RefCOCO+ / RefCOCOg | 80.1% | #27 of 29 | Lab model card | |
| OCRBench v2 | 43.7% | #21 of 22 | Lab model card | |
| MVBench | 58.2% | #22 of 23 | Lab model card | |
| AI2D | 75.4% | #50 of 52 | Lab model card | |
| ERQA | 42% | #35 of 36 | Lab model card | |
| MMMU-Pro | 26.4% | #90 of 90 | Lab model card | |
| LogicVista | 34.7% | #14 of 14 | Lab model card | |
| LingoQA | 58.2% | #10 of 10 | Lab model card | |
| RoboSpatial-Home | 63.4% | #3 of 3 | Lab model card | |
| CausalVQA | 37% | #2 of 2 | Lab model card | |
| VideoPhy-2 | 40.3% | #2 of 2 | Lab model card |
Every result in the longer documents the lab published with it. 23 of them aren't in the launch post.