MiniCPM-V: benchmark results
The original MiniCPM-V (also described as OmniLMM-3B), built from SigLIP-400M and MiniCPM-2.4B connected by a perceiver resampler. The language component size is not the full multimodal parameter total. Provider: OpenBMB. Released 2024-02-01. Access: Open.
Aggregate Rating 1296 ± 33, rank #8797 of 10916 rated models, from 411 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| OpenVLM MMT-Bench - MeViS | 50 | Score (%) | 96.1 |
| OpenVLM MMT-Bench - Texture Material Recognition | 85 | Score (%) | 95.4 |
| OpenVLM MMT-Bench - VI | 66.6 | Score (%) | 95.1 |
| OpenVLM MMT-Bench - Color Constancy | 71.4 | Score (%) | 94.2 |
| OpenVLM COCO Captions | 63.3 | CIDEr | 93.3 |
| OpenVLM MMT-Bench - Other Biological Attributes | 85 | Score (%) | 93.2 |
| OpenVLM MMT-Bench - Color Assimilation | 75 | Score (%) | 93 |
| OpenVLM MMT-Bench - Religious Recognition | 75 | Score (%) | 89.6 |
| OpenVLM MMT-Bench - Multiple View Image Understanding | 40 | Score (%) | 89.3 |
| OpenVLM MMT-Bench - Scene Recognition | 80 | Score (%) | 88.1 |
| OpenVLM MME - Celebrity | 163.5 | Score | 87.2 |
| OpenVLM MMT-Bench - Animated Character Recognition | 60 | Score (%) | 85 |
Interactive version: theaggregate.ai/model?slug=minicpm-v · How It Works · Data refreshed daily, snapshot 2026-11-08.