MMSI-Bench: leaderboard
Multi-image spatial intelligence benchmark testing VLMs on positional relationships, attribute reasoning, and motion reasoning across multiple images.
Metric: Accuracy (%). Source: github.com. Status: saturation imminent. 38 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Median Expert | 97.2 |
| 2 | Gemini 3 Pro | 49.2 |
| 3 | GPT-5 | 41.9 |
| 4 | O3 | 41 |
| 5 | GPT-4.5 | 40.3 |
| 6 | Gemini 2.5 Pro (Thinking) | 37 |
| 7 | Gemini 2.5 Pro | 36.9 |
| 8 | Doubao-1.5-Pro | 33 |
| 9 | GPT-4.1 | 30.9 |
| 10 | GPT-4o | 30.3 |
| 11 | Claude 3.7 Sonnet (Thinking) | 30.2 |
| 12 | InternVL2.5-2B | 29 |
| 13 | InternVL3-78B | 28.5 |
| 14 | InternVL2.5-78B | 28.5 |
| 15 | InternVL3-8B | 25.7 |
Interactive version: theaggregate.ai/benchmark?slug=mmsi-bench · How It Works · Data refreshed daily, snapshot 2026-09-05.