PRISM-VLM - Quality: leaderboard
The quality axis of PRISM-VLM (NAVER, 2026): accuracy of compact vision-language models on the original, unperturbed questions of 6,238 items recycled from 15 public VLM benchmarks, graded against the source answers by a GPT-5 judge. Score in percent; higher is better.
Metric: Per-axis mean score (%; base accuracy on the recycled question, judged against the source gold answer; 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 July 2027. 42 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Minimal) | 74.1 | #159 (Gemini 3 Flash) |
| 2 | Gemini 3.1 Flash Lite (Minimal) | 71.3 | #363 (Gemini 3.1 Flash Lite) |
| 3 | Qwen 3.5 9B (Non-reasoning) | 69.9 | #803 (Qwen 3.5 9B) |
| 4 | GPT-4.1 Mini | 69.7 | #775 |
| 5 | GPT-5.4 Mini | 69.2 | #389 |
| 6 | Gemini 2.5 Flash (Non-reasoning) | 68.9 | #452 (Gemini 2.5 Flash) |
| 7 | Gemini 2.0 Flash | 68.6 | #702 |
| 8 | Qwen 3.5 4B (Non-reasoning) | 68.2 | #1180 (Qwen 3.5 4B) |
| 9 | Claude Haiku 4.5 | 66.8 | #568 |
| 10 | Qwen 3 VL 8B Instruct | 66.5 | #982 |
| 11 | Gemini 2.0 Flash Lite | 66.5 | #1032 |
| 12 | GPT-5 Mini (Minimal) | 66.4 | #349 (GPT-5 Mini) |
| 13 | Qwen 3 VL 4B Instruct | 66.1 | #1281 |
| 14 | Nova 2.0 Lite | 66.1 | #1133 |
| 15 | Gemini 2.5 Flash Lite | 64.6 | #958 |
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-quality · How It Works · Data refreshed daily, snapshot 2026-11-12.