MMAlignBench: leaderboard
MMAlignBench (OmniAlign-V) — human-preference alignment of multimodal model responses to open-ended image questions; judged against a GPT-4o reference, reward score (-100 to 100).
Metric: Overall Reward. Source: huggingface.co. Saturation forecast: Around January 2027. 29 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Claude 3.7 Sonnet | 60.32 | #249 |
| 2 | GPT-4.1 (2025-04-14) | 59.92 | #250 |
| 3 | Grok 2 (1212) | 53.17 | #496 |
| 4 | Gemini 2.0 Flash | 40.28 | #342 |
| 5 | Claude 3.5 Sonnet (20241022) | 34.92 | #291 |
| 6 | Gemini 2.5 Flash | 32.14 | #245 |
| 7 | Pixtral-12B | 21.23 | #808 |
| 8 | Llama 4 Scout Instruct | 14.29 | #539 |
| 9 | Qwen 2.5 VL 7B | 7.54 | #643 |
| 10 | InternVL3-78B | -0.79 | #352 |
| 11 | InternVL3-8B | -7.94 | #620 |
| 12 | InternVL3-38B | -14.68 | #405 |
| 13 | InternVL2.5-78B | -19.64 | #431 |
| 14 | Llama 3.2 11B Instruct | -37.7 | #1126 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=mmalignbench · How It Works · Data refreshed daily, snapshot 2026-10-20.