RefSpatialBench: leaderboard
Reference-based spatial benchmark evaluating whether multimodal models understand relative positions and spatial language in images.
Metric: RefSpatialBench (imported). Source: huggingface.co. Source: benchmarklist.com. Source: llm-stats.com. Saturation forecast: Around August 2027. 4 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.6 27B | 70 | #170 |
| 2 | Qwen 3 VL 235B A22B (Thinking) | 69.9 | #238 (Qwen 3 VL 235B A22B) |
| 3 | Qwen 3.5 122B A10B | 69.3 | #173 |
| 4 | Qwen 3.5 27B | 67.7 | #187 |
| 5 | Qwen 3.6 35B A3B | 64.3 | #226 |
| 6 | Qwen 3.5 35B A3B | 63.5 | #256 |
| 7 | Seed 1.8 | 56.3 | #139 |
| 8 | Seed 2.0 Mini | 55.6 | #276 |
| 9 | Gemma 4 31B | 4.7 | #200 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=refspatialbench · How It Works · Data refreshed daily, snapshot 2026-10-20.