BunchCount - Individual Counting: leaderboard
BunchCount (Wang et al., 2026) tests object counting in real images whose objects also form semantic groups (bunches of grapes, stacks of plates, pairs of shoes). This board asks vision-language models, zero-shot, for the number of individual instances in the 310 test images of 14 categories unseen in training, about 90 per image; mean absolute error, lower is better.
Metric: Mean absolute error (individual instances per image, lower is better; individual-level test split, the 310 real images of 14 categories unseen in training, about 90 instances per image; vision-language models zero-shot with reasoning disabled, temperature 0, one-integer answer). Source: arxiv.org. Saturation forecast: Around August 2027. 8 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.6 Sol (Non-reasoning) | 15.94 | #30 (GPT-5.6 Sol) |
| 2 | Claude Opus 5 (Non-reasoning) | 22.05 | #22 (Claude Opus 5) |
| 3 | Qwen 3.7 Max (Non-reasoning) | 31.02 | #145 (Qwen 3.7 Max) |
| 4 | Qwen 3.5 27B (Non-reasoning) | 38.28 | #359 (Qwen 3.5 27B) |
| 5 | Qwen 3 VL 8B Instruct | 40.07 | #981 |
| 6 | GLM-4.5V (Non-reasoning) | 44.5 | #744 (GLM-4.5V) |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=bunchcount-individual-counting · How It Works · Data refreshed daily, snapshot 2026-11-13.