SpaCE-10: leaderboard

Each model answers the single-answer multiple-choice questions of SpaCE-10 about real indoor 3D scans, given 8 video frames; the questions combine several spatial skills (counting, size, relative position, situated observation, planning) in one item. The score is the paper's task-weighted overall accuracy, with a GPT-4o judge mapping answers to options; unlike single-skill spatial suites such as VSI-Bench, every question composes several atomic skills.

Metric: Accuracy (%). Source: arxiv.org. Saturation forecast: Around August 2027. 34 models tracked. Updated 17 Nov 2026.

Top models

#ModelScoreOverall rank
1InternVL3.5-241B-A28B55#720
2InternVL3-38B53.1#927
3InternVL3-78B52.5#755
4GLM-4.5V51.6#758
5InternVL3.5-20B-A4B51.6
6InternVL3.5-14B48.8#1503
7Qwen 3 VL 235B A22B Instruct47.9#540
8InternVL3-14B47.3#1191
9InternVL2.5-78B47.1#1036
10Claude 3.7 Sonnet46.2#476
11InternVL2.5-38B45.1#1366
12InternVL3-2B44.2#3323
13InternVL3.5-38B43.9#1017
14InternVL2.5-26B43.3#2191
15InternVL2.5-8B41.8#2401

No result here: #1 Claude Fable 5.2, #2 GPT-6.2 Astra, #3 Gemini 4 Carbon.

Interactive version: theaggregate.ai/benchmark?slug=space-10 · How It Works · Data refreshed daily, snapshot 2026-11-17.