OpenVLM Subjective: leaderboard
OpenVLM Subjective Leaderboard (VLMEvalKit/OpenCompass) — aggregate of subjective vision-language evaluations (Creation-MMBench, MIA-Bench, MM-IFEval, MMAlignBench, MM-Vet, WildVision), with pairwise rewards normalized to 0-100; average normalized score.
Metric: Average Normalized Score (%). Source: huggingface.co. Saturation forecast: Estimated already saturated. 29 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-4.1 (2025-04-14) | 76.6 | #250 |
| 2 | Gemini 2.5 Flash | 71.73 | #245 |
| 3 | Claude 3.7 Sonnet | 71.47 | #249 |
| 4 | Gemini 2.0 Flash | 70.11 | #342 |
| 5 | Grok 2 (1212) | 68.91 | #496 |
| 6 | Claude 3.5 Sonnet (20241022) | 68.9 | #291 |
| 7 | InternVL3-78B | 67.61 | #352 |
| 8 | Llama 4 Scout Instruct | 65.54 | #539 |
| 9 | InternVL3-38B | 64.31 | #405 |
| 10 | InternVL3-8B | 63.66 | #620 |
| 11 | InternVL2.5-78B | 62.7 | #431 |
| 12 | Pixtral-12B | 60.42 | #808 |
| 13 | Qwen 2.5 VL 7B | 59.66 | #643 |
| 14 | Llama 3.2 11B Instruct | 48.1 | #1126 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=openvlm-subjective · How It Works · Data refreshed daily, snapshot 2026-10-20.