MiniCPM-o-2.6: benchmark results
An end-to-end multimodal checkpoint combining SigLIP-400M, Whisper-medium-300M, ChatTTS-200M and Qwen2.5-7B, with a publisher-rounded total of 8B parameters. Speech and vision modules are part of the model rather than language-only parameters. Provider: OpenBMB. Released 2025-01-13. Access: Open.
Aggregate Rating 1496 ± 10, rank #2415 of 10916 rated models, from 793 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| OpenVLM Q-Bench - How (In-context Distortion) | 73.6 | Accuracy (%) | 99.1 |
| OpenVLM AI2D - Eclipses | 100 | Accuracy (%) | 98.6 |
| OpenVLM Q-Bench - Yes-or-No (In-context Others) | 92.9 | Accuracy (%) | 98.4 |
| OpenVLM MMT-Bench - Musical Instrument Recognition | 100 | Score (%) | 98.3 |
| OpenVLM OCRBench - Key Information Extraction | 93.5 | Correct Answers (%) | 97.9 |
| OpenVLM ScienceQA Test | 98.8 | Accuracy (%) | 97.7 |
| OpenVLM MMBench V1.0 CN - AR | 87.8 | Accuracy (%) | 97.6 |
| OpenVLM Q-Bench - How (Others) | 77.9 | Accuracy (%) | 97.5 |
| OpenVLM MMT-Bench - Face Retrieval | 100 | Score (%) | 97.3 |
| OpenVLM MMT-Bench - Traffic Sign Understanding | 95 | Score (%) | 97.3 |
| OpenVLM OCRBench - Doc-oriented VQA | 91 | Correct Answers (%) | 97.2 |
| OpenVLM MMT-Bench - Abstract Visual Recognition | 90 | Score (%) | 96.8 |
Interactive version: theaggregate.ai/model?slug=minicpm-o-2-6 · How It Works · Data refreshed daily, snapshot 2026-11-08.