MiniCPM-V-4.5: benchmark results
A multimodal checkpoint built on Qwen3-8B and SigLIP2-400M, with a publisher-rounded total of 8B parameters. It supports thinking and non-thinking modes; the catalog keeps their evaluated operating modes separate. Provider: OpenBMB. Released 2025-08-26. Access: Open.
Aggregate Rating 1566 ± 9, rank #1451 of 10916 rated models, from 426 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| Holtercare-Bench - Event Counting (Text) | 51.38 | Accuracy (%; multiple-choice closed QA on how many times a g | 100 |
| MVPBench - Olympic Sports Matching | 37.14 | MVPBench normalized proficiency score (accuracy minus chance | 100 |
| MoHallBench - Visual Similarity Hallucination (Multiple Choice) | 64.1 | Multiple-choice accuracy (%; dual-video four-option choice w | 100 |
| RoadSafe365 - Area | 79.76 | Four-option multiple-choice accuracy (%) on the area (urban, | 100 |
| MathVista | 80.2 | Accuracy (%) | 96.5 |
| PRISM-VLM - Sycophancy Resistance | 79.1 | Per-axis mean score (%; share of turn-0-correct answers held | 95.1 |
| EgoIntrospect - Proactive Timing Judgment | 60 | Macro-F1 (%) over the interruptible and do-not-disturb class | 95 |
| HallusionBench | 61.3 | Overall Score | 93.7 |
| CSR-Bench (Hallucination) | 42.3 | Alignment accuracy (%) averaged over the four question forms | 93.3 |
| CSR-Bench (Hallucination) - Modality Conflict | 28.5 | Alignment accuracy (%) on a structurally similar question wh | 93.3 |
| MMStar | 71.7 | Accuracy (%) | 91.7 |
| VNU-Bench - Event Comparison | 53.45 | Accuracy (%; question type T2, 4-option multiple choice) | 91.7 |
Interactive version: theaggregate.ai/model?slug=minicpm-v-4-5 · How It Works · Data refreshed daily, snapshot 2026-11-08.