MIA-Bench: leaderboard
MIA-Bench (Apple) — 400 image-prompt pairs with layered, compositional instructions testing how strictly multimodal models adhere to complex instructions; judge-scored overall score.
Metric: Overall 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) | 94.55 | #250 |
| 2 | Grok 2 (1212) | 93.61 | #496 |
| 3 | Gemini 2.0 Flash | 92.93 | #342 |
| 4 | InternVL3-78B | 92.65 | #352 |
| 5 | Gemini 2.5 Flash | 92.17 | #245 |
| 6 | InternVL3-38B | 92.05 | #405 |
| 7 | InternVL2.5-78B | 91.97 | #431 |
| 8 | Llama 4 Scout Instruct | 91.57 | #539 |
| 9 | Claude 3.5 Sonnet (20241022) | 91.46 | #291 |
| 10 | Pixtral-12B | 90.28 | #808 |
| 11 | Claude 3.7 Sonnet | 89.9 | #249 |
| 12 | InternVL3-8B | 89.7 | #620 |
| 13 | Qwen 2.5 VL 7B | 85.9 | #643 |
| 14 | Llama 3.2 11B Instruct | 76.57 | #1126 |
No result here: #3 Claude Opus 5.5, #6 GPT-6 Astra, #7 GPT-6.1 Sol.
Interactive version: theaggregate.ai/benchmark?slug=mia-bench · How It Works · Data refreshed daily, snapshot 2026-10-20.