Molmo2-8B: benchmark results
Allen AI's Molmo2-8B, an open vision-language model built on Qwen3-8B with a SigLIP 2 vision backbone, trained for image, video and multi-image understanding and grounding. Provider: Allen AI. Released 2025-12-11. Access: Open.
Aggregate Rating 1570 ± 9, rank #543 of 1656 rated models, from 197 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| EgoMemReason - Temporal Counting | 50 | Accuracy (%; 100 multiple-choice questions) | 100 |
| MoHallBench - Object-Action Hallucination (Multiple Choice) | 75 | Multiple-choice accuracy (%; four-option multiple choice (gr | 100 |
| GQA | 87.07 | Overall Accuracy (%) | 99.8 |
| CringeBench | 0.63 | Cringe Score (0-10, lower is better) | 90.9 |
| VLM-RobustBench (MMMU-Pro) | 1 | Mean relative corruption error (percent of visual gain lost) | 87.5 |
| VSTAT - Atomic | 34.7 | Score (%) on questions about atomic state structures: 1,500 | 86.4 |
| Molmo2-CapTest | 43.2 | Caption F1 (%; LLM-judged statement precision and recall aga | 85 |
| Molmo2-VideoCount | 35.5 | Exact-count accuracy (%; 533 validation queries) | 85 |
| TimeBlind | 31.2 | Instance accuracy I-Acc (%) over TimeBlind's 600 instances ( | 81 |
| TimeBlind - Accuracy | 68.7 | Standard accuracy (%) over all 2,400 TimeBlind video-questio | 81 |
| BURST-VideoCount | 60.8 | Exact-count accuracy (%; 2.2k queries from BURST test-set tr | 80 |
| Molmo2 Human Preference | 1057 | Elo rating (bootstrap median, Bradley-Terry; pairwise human | 80 |
Interactive version: theaggregate.ai/model?slug=molmo2-8b · How It Works · Data refreshed daily, snapshot 2026-10-20.