MiniCPM-V-4: benchmark results
A multimodal checkpoint combining SigLIP2-400M with MiniCPM4-3B. The publisher reports 4.1B total parameters including the model components and supports single-image, multi-image and video understanding. Provider: OpenBMB. Released 2025-08-02. Access: Open.
Aggregate Rating 1521 ± 29, rank #2060 of 10916 rated models, from 35 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| FREAK (Multiple Choice) | 46.06 | Accuracy (%) on FREAK's 1,000 multiple-choice questions, eac | 100 |
| FREAK - Counting | 24.91 | Accuracy (%) on the counting questions of FREAK's 1,799 ques | 100 |
| FREAK - Position | 40.88 | Accuracy (%) on the position questions of FREAK's 1,799 ques | 94.1 |
| OpenVLM OCRBench | 89.7 | Score (normalized) | 93.7 |
| FREAK (Multiple Choice) - Consistency Accuracy | 34.5 | Consistency accuracy (%): share of FREAK's 1,000 multiple-ch | 90.9 |
| FREAK - Attribute | 48.97 | Accuracy (%) on the attribute questions of FREAK's 1,799 que | 88.2 |
| HallusionBench | 52.6 | Overall Score | 75.1 |
| MM-Vet | 66.5 | Score | 74.3 |
| MathVista | 67.5 | Accuracy (%) | 72.3 |
| FREAK - Analysis | 30.4 | Accuracy (%) on the analysis questions of FREAK's 1,799 ques | 70.6 |
| FREAK - Detection | 46.12 | Accuracy (%) on the detection questions of FREAK's 1,799 que | 70.6 |
| MMStar | 61.9 | Accuracy (%) | 66.8 |
Interactive version: theaggregate.ai/model?slug=minicpm-v-4 · How It Works · Data refreshed daily, snapshot 2026-11-08.