OpenRef - Ordinal: leaderboard

The ordinal split of OpenRef (2026), an open-world referring expression benchmark: expressions with ordinal terms, to be grounded with bounding boxes in the image. The score is F1 over predicted and ground-truth boxes in percent; higher is better.

Metric: Grounding F1 (%) on expressions with ordinal terms, zero-shot box grounding, each model prompted in its own grounding format; higher is better. Source: arxiv.org. Saturation forecast: Around May 2027. 13 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 2.5 VL 32B46.7#941
2InternVL3.5-8B35.4#1604
3Qwen 3 VL 8B27.7#917
4MiniCPM-V-4.524.6#1441
5GLM-4.6V-Flash24.1#1040
6Keye-VL-1.5-8B23.6#1720
7Qwen 2.5 VL 7B15.9#1862
8Qwen 3 VL 4B15.9#1151
9MiMo-VL-7B-RL15.1#1155
10Qwen 3 VL 2B12.3#2134
11Qwen 2 VL 7B9.2#2275
12LLaVA-OneVision-1.5-8B6.7#1864
13OpenRef Mistral-3 8B (checkpoint unspecified)2.3

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

Interactive version: theaggregate.ai/benchmark?slug=openref-ordinal · How It Works · Data refreshed daily, snapshot 2026-11-21.