SmolVLM2-2.2B-Instruct: benchmark results
Provider: HuggingFace. Released 2025-02-20. Access: Open.
Aggregate Rating 1347 ± 36, rank #7791 of 10916 rated models, from 113 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| OpenVLM Video - Video-MME (w/o subs) - Short Videos | 79 | Accuracy (%) | 86.1 |
| ProactiveBench - ChangeIt | 25.5 | Trajectory accuracy (%) on the ambiguous ChangeIt video fram | 66.7 |
| ProactiveBench - MVP-N | 11.1 | Trajectory accuracy (%) on the uninformative multi-view obje | 66.7 |
| OpenVLM Video - MMBench-Video - Coarse Perception | 1.7 | GPT-4 Rating (0-3) | 55.7 |
| OpenVLM Video - MMBench-Video - Relation Reasoning | 1.48 | GPT-4 Rating (0-3) | 53.4 |
| OpenVLM Video - MMBench-Video - Temporal Reasoning | 1.24 | GPT-4 Rating (0-3) | 51.1 |
| OpenVLM Video - MMBench-Video - Commonsense Reasoning | 1.46 | GPT-4 Rating (0-3) | 50 |
| OpenVLM Video - MMBench-Video - Single-Instance Fine-grained Perception | 1.5 | GPT-4 Rating (0-3) | 50 |
| OpenVLM Video - MVBench - Object Interaction | 70 | Accuracy (%) | 50 |
| ProactiveBench - VSOD | 11.9 | Trajectory accuracy (%) on the temporally occluded video fra | 47.6 |
| OpenVLM Video - MVBench - Scene Transition | 89 | Accuracy (%) | 47.2 |
| OpenVLM Video - MMBench-Video - Cross-Instance Fine-grained Perception | 1.18 | GPT-4 Rating (0-3) | 46.6 |
Interactive version: theaggregate.ai/model?slug=smolvlm2-2-2b-instruct · How It Works · Data refreshed daily, snapshot 2026-11-08.