EPIC-Bench - Basic Attributes: leaderboard
Metric: Target localization score (0-100) on the basic-attribute tasks (exhaustive grounding of every object matching a colour, material, shape or category description), on EPIC-Bench (6,661 human-annotated image, text and mask tuples from 25 public datasets across 23 fine-grained embodied perception tasks; the model outputs bounding boxes, counts, path points or feasibility judgements that are scored against the masks), zero-shot, averaged over six runs for local models and two for API models; higher is better. Source: arxiv.org. Saturation forecast: Around August 2028. 89 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Qwen 3 VL 235B A22B (Thinking) | 58 |
| 2 | GPT-5.5 (Non-reasoning) | 57.78 |
| 3 | Gemini 3 Pro | 56.81 |
| 4 | Qwen 3.5 397B A17B | 55.45 |
| 5 | Gemini 3.1 Pro (Preview) | 55.38 |
| 6 | Seed 1.8 | 55.29 |
| 7 | Qwen 3.5 Plus | 55.09 |
| 8 | Qwen 3.5 27B | 54.15 |
| 9 | Qwen 3.5 122B A10B | 53.61 |
| 10 | Qwen 3.5 35B A3B | 53.51 |
| 11 | Qwen 3 VL 30B A3B (Thinking) | 52.69 |
| 12 | Qwen 3.6 27B (Non-reasoning) | 51.04 |
| 13 | Qwen 3.5 122B A10B (Non-reasoning) | 50.77 |
| 14 | Qwen 3.5 397B A17B (Non-reasoning) | 50.56 |
| 15 | Gemini 3 Flash (Preview) | 50.26 |
Interactive version: theaggregate.ai/benchmark?slug=epic-bench-basic-attributes · How It Works · Data refreshed daily, snapshot 2026-10-07.