BunchCount - Group Counting: leaderboard
BunchCount (2026) asks models to count semantic groups formed by several objects, such as bunches of grapes, stacks of plates or pairs of shoes, rather than the individual objects. Mean absolute error of the predicted group count on the 310-image test split, whose object categories do not occur in training; vision-language models answer zero-shot with reasoning disabled. Lower is better.
Metric: Mean absolute error (semantic groups per image, lower is better; group-level test split of 310 real images with 14 categories unseen in training, e.g. bunches of grapes, stacks of plates, pairs of shoes; 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) | 5.46 | #29 (GPT-5.6 Sol) |
| 2 | Qwen 3.7 Max (Non-reasoning) | 5.55 | #142 (Qwen 3.7 Max) |
| 3 | Qwen 3.5 27B (Non-reasoning) | 5.84 | #358 (Qwen 3.5 27B) |
| 4 | GLM-4.5V (Non-reasoning) | 8.83 | #739 (GLM-4.5V) |
| 5 | Claude Opus 5 (Non-reasoning) | 10.26 | #20 (Claude Opus 5) |
| 6 | Qwen 3 VL 8B Instruct | 10.67 | #982 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=bunchcount-group-counting · How It Works · Data refreshed daily, snapshot 2026-11-12.