MiniCPM-V-2_6

MiniCPM-V 2.6 is an 8-billion-parameter open-source vision-language model from OpenBMB, built on a SigLip-400M vision encoder and Qwen2-7B language backbone. Released in August 2024, it targets efficient on-device deployment while delivering GPT-4V-class performance across single-image, multi-image, and video understanding tasks. The model supports arbitrary aspect-ratio images up to 1.8 million pixels and achieves state-of-the-art token density, encoding a full 1.8MP image into only 640 visual tokens.

On standard multimodal benchmarks, MiniCPM-V 2.6 scores 65.2 on OpenCompass (averaging 8 popular vision benchmarks), 49.8 on MMMU, 60.6 on MathVista, 82.1 on AI2D, 90.8 on DocVQA, and 852 on OCRBench—surpassing proprietary models including GPT-4o mini, GPT-4V, Gemini 1.5 Pro, and Claude 3.5 Sonnet on single-image understanding despite its compact size. It also leads on multi-image benchmarks (Mantis-Eval 69.1, Mathverse mv 84.9) and video benchmarks (Video-MME 60.9 without subtitles, 63.6 with subtitles).

The model incorporates RLAIF-V and VisCPM techniques for reduced hallucination, strong multilingual OCR in 30+ languages, and flexible deployment via HuggingFace Transformers, vLLM, llama.cpp, Ollama, and INT4/GGUF quantized variants. It is free for academic research and available for commercial use after registration under the MiniCPM Model License.