RefSpatialBench: leaderboard
Reference-based spatial benchmark evaluating whether multimodal models understand relative positions and spatial language in images.
Metric: RefSpatialBench (self-reported). Source: benchmarklist.com. Status: saturation imminent. 4 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Qwen 3.6 27B | 70 |
| 2 | Qwen 3 VL 235B A22B (Thinking) | 69.9 |
| 3 | Qwen 3.5 122B A10B | 69.3 |
| 4 | Qwen 3.5 27B | 67.7 |
| 5 | Qwen 3.6 35B A3B | 64.3 |
| 6 | Qwen 3.5 35B A3B | 63.5 |
| 7 | Gemma 4 31B | 4.7 |
Interactive version: theaggregate.ai/benchmark?slug=refspatialbench · How It Works · Data refreshed daily, snapshot 2026-09-05.