Abstract

We introduce InternVL 2.5, an advanced multimodal large language model (MLLM)series that builds upon InternVL 2.0, maintaining its core model architecturewhile introducing significant enhancements in training and testing strategiesas well as data quality. In this work, we delve into the relationship betweenmodel scaling and performance, systematically exploring the performance trendsin vision encoders, language models, dataset sizes, and test-timeconfigurations. Through extensive evaluations on a wide range of benchmarks,including multi-discipline reasoning, document understanding, multi-image /video understanding, real-world comprehension, multimodal hallucinationdetection, visual grounding, multilingual capabilities, and pure languageprocessing, InternVL 2.5 exhibits competitive performance, rivaling leadingcommercial models such as GPT-4o and Claude-3.5-Sonnet. Notably, our model isthe first open-source MLLMs to surpass 70% on the MMMU benchmark, achieving a3.7-point improvement through Chain-of-Thought (CoT) reasoning and showcasingstrong potential for test-time scaling. We hope this model contributes to theopen-source community by setting new standards for developing and applyingmultimodal AI systems. HuggingFace demo seehttps://huggingface.co/spaces/OpenGVLab/InternVL