OpenLMM Spatial - Spatial457: leaderboard
Spatial457 accuracy from the OpenCompass OpenLMM Spatial Leaderboard (VLMEvalKit): open-ended questions on synthetic scenes across five levels, from single-object attributes to 3D pose and collision reasoning. Accuracy in percent, higher is better.
Metric: Accuracy (%). Source: huggingface.co. Saturation forecast: Around January 2028. 57 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 2 VL 72B | 67.4 | #450 |
| 2 | InternVL3-38B | 64.3 | #413 |
| 3 | Qwen 2 VL 7B | 62.6 | #743 |
| 4 | GPT-5 | 62.3 | #105 |
| 5 | InternVL3-14B | 60.9 | #506 |
| 6 | InternVL3-78B | 60.4 | #361 |
| 7 | GPT-5 Mini (2025-08-07) | 58 | #175 |
| 8 | GPT-4.1 (2025-04-14) | 57.2 | #257 |
| 9 | InternVL3-8B | 56.6 | #624 |
| 10 | Gemini 2.0 Flash | 53.2 | #351 |
| 11 | Gemini 1.5 Pro (002) | 53.1 | #368 |
| 12 | Gemini 2.5 Flash | 52.9 | #255 |
| 13 | GPT-4.1 Mini | 51.1 | #363 |
| 14 | Gemini 1.5 Flash (002) | 50.9 | #587 |
| 15 | Claude 3.7 Sonnet | 49.3 | #259 |
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-spatial457 · How It Works · Data refreshed daily, snapshot 2026-11-05.