Phi-4 Multimodal: benchmark results
Provider: Microsoft. Released 2025-02-26. Access: Open.
Aggregate Rating 1441 ± 14, rank #4159 of 10916 rated models, from 1045 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| TRAP - Image Privacy | 44 | Privacy score (%, 1 minus leakage) on the 200 image-document | 100 |
| TRAP - Multimodal Privacy | 48.5 | Privacy score (%, 1 minus leakage) on the 200 multimodal-doc | 100 |
| TRAP - Privacy | 38.4 | Privacy score (%): 1 minus the leakage rate under the attack | 100 |
| OpenVLM MMT-Bench - Raven Progressive Matrices | 70 | Score (%) | 99.5 |
| OpenVLM A-Bench - Attribute Awareness | 96.4 | Accuracy (%) | 97.8 |
| MEGA-Bench Task - MMSoc HatefulMemes | 85.7 | Task Score (%) | 97.7 |
| OpenVLM MMT-Bench - Scene Recognition | 85 | Score (%) | 97.3 |
| OpenVLM MMT-Bench - Small Object Detection | 70 | Score (%) | 96.8 |
| MEGA-Bench Task - MMSoc Memotion | 68.2 | Task Score (%) | 96.5 |
| OpenVLM MMT-Bench - Rotated Object Detection | 77.8 | Score (%) | 96.4 |
| OpenVLM MMT-Bench - Camouflage Object Detection | 80 | Score (%) | 96.1 |
| OpenVLM MMT-Bench - Clock Reading | 80 | Score (%) | 95.6 |
Interactive version: theaggregate.ai/model?slug=phi-4-multimodal · How It Works · Data refreshed daily, snapshot 2026-11-08.