PRISM-VLM - Counting: leaderboard
The counting axis of PRISM-VLM (NAVER, 2026): synthesized questions ask how many of an object an image shows, with integer answers from 1 to 100 checked by exact match. Accuracy in percent; higher is better.
Metric: Per-axis mean score (%; exact-integer accuracy on synthesized counting questions; over PRISM-VLM's 6,238 items recycled from 15 public VLM benchmarks (five seeds of 100 items per benchmark); GPT-5 (low effort) synthesizes the perturbations and grades the open-ended axes; each model at the lowest reasoning effort its provider exposes, temperature 0). Source: arxiv.org. Saturation forecast: Around May 2028. 42 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.5 9B (Non-reasoning) | 66.4 | #803 (Qwen 3.5 9B) |
| 2 | Gemini 3 Flash (Minimal) | 65.1 | #159 (Gemini 3 Flash) |
| 3 | Qwen 3.5 4B (Non-reasoning) | 64.6 | #1180 (Qwen 3.5 4B) |
| 4 | Gemini 3.1 Flash Lite (Minimal) | 62.8 | #363 (Gemini 3.1 Flash Lite) |
| 5 | GPT-4.1 Mini | 61.8 | #775 |
| 6 | Qwen 3 VL 8B Instruct | 61.8 | #982 |
| 7 | GPT-5.4 Mini | 61.7 | #389 |
| 8 | Gemini 2.5 Flash (Non-reasoning) | 61.7 | #452 (Gemini 2.5 Flash) |
| 9 | GPT-5 Mini (Minimal) | 61.5 | #349 (GPT-5 Mini) |
| 10 | Qwen 3 VL 4B Instruct | 58.5 | #1281 |
| 11 | Qwen 3.5 2B (Non-reasoning) | 57.9 | #4849 (Qwen 3.5 2B) |
| 12 | Gemini 2.0 Flash | 57.3 | #702 |
| 13 | Nova 2.0 Lite | 56.2 | #1133 |
| 14 | Gemini 2.0 Flash Lite | 55.1 | #1032 |
| 15 | Gemma 4 E4B | 54.8 | #2217 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=prism-vlm-counting · How It Works · Data refreshed daily, snapshot 2026-11-12.