PRISM-VLM: leaderboard
PRISM-VLM (NAVER, 2026) evaluates compact vision-language models, open-weight models up to 10B parameters and low-cost API tiers, on 6,238 items recycled from 15 public VLM benchmarks and perturbed along seven axes. PScore is the harmonic mean of the seven per-axis mean scores in percent, so one weak axis pulls it down; higher is better. GPT-5 generates the perturbations and grades the open-ended axes, and each model runs at the lowest reasoning effort its provider exposes.
Metric: PScore (%; harmonic mean of the seven per-axis mean scores, x100, 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 February 2027. 42 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3 Flash (Minimal) | 71.9 | #159 (Gemini 3 Flash) |
| 2 | Gemini 3.1 Flash Lite (Minimal) | 67.5 | #363 (Gemini 3.1 Flash Lite) |
| 3 | Gemini 2.5 Flash (Non-reasoning) | 67.3 | #452 (Gemini 2.5 Flash) |
| 4 | Qwen 3 VL 8B Instruct | 67.1 | #982 |
| 5 | Qwen 3 VL 4B Instruct | 64.2 | #1281 |
| 6 | GPT-5 Mini (Minimal) | 63.3 | #349 (GPT-5 Mini) |
| 7 | Nova 2.0 Lite | 62.8 | #1133 |
| 8 | Gemini 2.5 Flash Lite | 62.1 | #958 |
| 9 | Qwen 3.5 9B (Non-reasoning) | 61.1 | #803 (Qwen 3.5 9B) |
| 10 | Qwen 3.5 4B (Non-reasoning) | 59.8 | #1180 (Qwen 3.5 4B) |
| 11 | Gemini 2.0 Flash Lite | 59.1 | #1032 |
| 12 | GPT-4.1 Mini | 57.6 | #775 |
| 13 | Gemini 2.0 Flash | 57 | #702 |
| 14 | Grok 4 Fast (Non-reasoning) | 55.6 | #438 (Grok 4 Fast) |
| 15 | Qwen 3.5 2B (Non-reasoning) | 54.9 | #4849 (Qwen 3.5 2B) |
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 · How It Works · Data refreshed daily, snapshot 2026-11-12.