EPIC-Bench: leaderboard
Metric: Average score (0-100) over all samples, each sample scored by its task's weighted mix of localization IoU, counting, path and feasibility scores, 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 2029. 89 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | Gemini 3 Pro | 54.81 |
| 2 | Gemini 3.1 Pro (Preview) | 54.72 |
| 3 | Qwen 3 VL 235B A22B (Thinking) | 50.93 |
| 4 | GPT-5.5 (Non-reasoning) | 50.16 |
| 5 | Qwen 3.5 397B A17B | 47.47 |
| 6 | Qwen 3.5 122B A10B | 47.24 |
| 7 | Qwen 3.5 Plus | 47.1 |
| 8 | Seed 1.8 | 46.32 |
| 9 | Qwen 3.5 27B | 45.67 |
| 10 | Qwen 3.5 35B A3B | 45.58 |
| 11 | Qwen 3.6 Plus | 45.48 |
| 12 | Qwen 3.6 27B (Non-reasoning) | 45.38 |
| 13 | Qwen 3.5 397B A17B (Non-reasoning) | 45.16 |
| 14 | Qwen 3.5 27B (Non-reasoning) | 44.71 |
| 15 | Gemini 3 Flash (Preview) | 44.69 |
Interactive version: theaggregate.ai/benchmark?slug=epic-bench · How It Works · Data refreshed daily, snapshot 2026-10-07.