OpenRef - Landmark: leaderboard

The landmark split of OpenRef (2026), an open-world referring expression benchmark: proper-noun landmark expressions, 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 proper-noun landmark expressions, zero-shot box grounding, each model prompted in its own grounding format; higher is better. Source: arxiv.org. Saturation forecast: Around February 2027. 13 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Qwen 3 VL 8B88.6#917
2Qwen 3 VL 4B87.1#1151
3GLM-4.6V-Flash86.5#1040
4Qwen 2.5 VL 32B85.5#941
5Qwen 2.5 VL 7B85#1862
6Qwen 3 VL 2B83.1#2134
7MiniCPM-V-4.578.8#1441
8InternVL3.5-8B72.5#1604
9LLaVA-OneVision-1.5-8B70.5#1864
10Qwen 2 VL 7B68.4#2275
11Keye-VL-1.5-8B67.9#1720
12MiMo-VL-7B-RL60.7#1155
13OpenRef Mistral-3 8B (checkpoint unspecified)8.2

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

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