MiniCPM-V-2: benchmark results
MiniCPM-V 2.0 combines SigLIP-400M and MiniCPM-2.4B with a perceiver resampler and multimodal RLHF alignment. Its card labels the model 2.8B, while the language-family 2.4B convention excludes embeddings; these component labels are not an independently verified full-checkpoint total. Provider: OpenBMB. Released 2024-04-12. Access: Open.
Aggregate Rating 1335 ± 29, rank #8073 of 10916 rated models, from 450 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| OpenVLM MMT-Bench - Color Constancy | 85.7 | Score (%) | 98.8 |
| OpenVLM MMT-Bench - VI | 69.1 | Score (%) | 97.6 |
| OpenVLM MMT-Bench - Traffic Sign Understanding | 95 | Score (%) | 97.3 |
| OpenVLM MMT-Bench - Geometrical Relativity | 75 | Score (%) | 96.1 |
| OpenVLM MMT-Bench - Color Assimilation | 75 | Score (%) | 93 |
| OpenVLM MMT-Bench - Temporal Ordering | 40 | Score (%) | 92 |
| OpenVLM MMT-Bench - Artwork Emotion Recognition | 60 | Score (%) | 91.7 |
| OpenVLM MMT-Bench - Painting Recognition | 90 | Score (%) | 90.8 |
| OpenVLM Q-Bench - How (In-context Distortion) | 69 | Accuracy (%) | 89.9 |
| OpenVLM MMT-Bench - Depth Estimation | 75 | Score (%) | 89.1 |
| OpenVLM COCO Captions | 26.4 | CIDEr | 86.7 |
| OpenVLM MMT-Bench - Traffic Light Understanding | 95 | Score (%) | 85 |
Interactive version: theaggregate.ai/model?slug=minicpm-v-2 · How It Works · Data refreshed daily, snapshot 2026-11-08.