Benchmark Almanac

Cosmos3-Nano

Cosmos3-Nano ranks first on 2 of the 23 benchmarks with five or more models scored.

NVIDIA Released May 31st, 2026 Compare with others
29
Benchmarks with a score
2
Ranked first, out of five or more models
3
In the top three
NVIDIA May 2026

Cosmos3-Nano

Open weights

No Epoch Capabilities Index score yet
  • MMMU-Pro 41.1% #84 /90
  • IFBench 32% #75 /80
  • ERQA 44% #33 /36

Where it ranks

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.

Knowledge and language

BenchmarkScoreRankSource
HallusionBench 50.5% #29 of 31 Lab model card
IFBench 32% #75 of 80 Lab model card

Vision, audio and robotics

BenchmarkScoreRankSource
PAI-Bench 83.7% #1 of 6 Official leaderboard
PhAIL DROID: Everyday tasks 68.1% #1 of 6 Official leaderboard
RBench 58.4% #3 of 12 Official leaderboard
Physics-IQ 57.7% #1 of 4 Official leaderboard
Where2Place 64% #1 of 4 Lab model card
CVBench 86.5% #5 of 15 Lab model card
Physics-IQ Verified -3.00 #4 of 10 Official leaderboard
InfographicVQA 81.8% #11 of 26 Lab model card
DocVQA 94.2% #17 of 36 Lab model card
MVPBench 66.9% #2 of 4 Lab model card
CausalVQA 70% #1 of 2 Lab model card
VideoPhy-2 45.6% #1 of 2 Lab model card
MMBench 85.1% #6 of 11 Lab model card
MVBench 73.2% #15 of 23 Lab model card
RoboSpatial-Home 66.3% #2 of 3 Lab model card
OCRBench v2 60.1% #15 of 22 Lab model card
VSI-Bench 54.9% #5 of 7 Lab model card
AI2D 85% #38 of 52 Lab model card
RefSpatial-Bench 53.1% #12 of 16 Lab model card
RealWorldQA 72.2% #29 of 38 Lab model card
CountBench 84.8% #23 of 28 Lab model card
LogicVista 41.8% #12 of 14 Lab model card
RefCOCO / RefCOCO+ / RefCOCOg 84.3% #26 of 29 Lab model card
LingoQA 68.4% #9 of 10 Lab model card
ERQA 44% #33 of 36 Lab model card
MMMU-Pro 41.1% #84 of 90 Lab model card
WBench 74.2 #6 of 6 Official leaderboard

Its lab's numbers next to 2 leaderboards

Where the benchmark's own leaderboard, or Epoch AI's run, also has this model. Gaps mostly come from different harnesses, effort levels and task subsets. Every pair.

Benchmark Lab says Board says Gap
PAI-Bench 79.4% Cosmos3-Nano text-to-video PAIBench-G Overall, Table 12. Also Cosmos3-Nano image-to-video PAIBench-G Overall, Table 12 82.7 83.7% Leaderboard −4.3 pts
RBench 58.4% Cosmos3-Nano image-to-video RBench score, Table 12 58.4% Leaderboard, GPT judge Same

Its system card

Every result in the longer documents the lab published with it. 27 of them aren't in the launch post.