OpenLMM Spatial - 3DSRBench (Circular): leaderboard
3DSRBench from the OpenCompass OpenLMM Spatial Leaderboard (VLMEvalKit): 3D spatial reasoning questions on height, location, orientation and multi-object relations, scored with CircularEval plus FlipEval (every option rotation and the mirrored image must be answered correctly). Accuracy in percent, higher is better.
Metric: CircularEval + FlipEval Accuracy (%). Source: huggingface.co. Saturation forecast: Around August 2028. 58 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5 | 53.5 | #105 |
| 2 | InternVL3-78B | 53.3 | #361 |
| 3 | Gemini 2.5 Pro | 51.1 | #159 |
| 4 | InternVL3-38B | 50.7 | #413 |
| 5 | GPT-5 Mini (2025-08-07) | 50 | #175 |
| 6 | Seed 1.6 Flash | 48.5 | #438 |
| 7 | InternVL3-14B | 48.3 | #506 |
| 8 | InternVL2.5-78B | 47.9 | #434 |
| 9 | GPT-4.1 Mini | 47.2 | #363 |
| 10 | GPT-4.1 (2025-04-14) | 47.2 | #257 |
| 11 | InternVL3-8B | 46.5 | #624 |
| 12 | Qwen 2 VL 72B | 46.3 | #450 |
| 13 | Gemini 2.5 Flash | 45.6 | #255 |
| 14 | GPT-5 Nano | 43.6 | #439 |
| 15 | Claude Sonnet 4 | 42.7 | #207 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=openlmm-spatial-3dsrbench-circular · How It Works · Data refreshed daily, snapshot 2026-11-05.