.png)
April 16, 2025
Development in the field of multimodal large language models (MLLMs) is progressing rapidly. Commercial providers often dominate with powerful, yet proprietary models. However, the open-source community is continuously working to close this gap. A remarkable example of this is InternVL3, a new open-source MLLM introduced by OpenGVLab.
InternVL3 has achieved a considerable score of 72.2 points in the MMMU benchmark, an established test procedure for evaluating multimodal models. This sets a new standard among freely available MLLMs and moves closer to the performance of commercial models. This progress is an important step towards the democratization of AI technologies, as it allows researchers, developers, and companies access to powerful multimodal models without relying on proprietary solutions.
Multimodal models are capable of processing and understanding different data types such as text, images, and audio simultaneously. This opens up a variety of application possibilities that go beyond the capabilities of purely text-based AI systems. Examples include image captioning, answering questions about images, generating images from text descriptions, and translation between different modalities.
The increasing performance of MLLMs like InternVL3 is driving the development of innovative applications in various fields, including:
- Robotics - Medical Diagnostics - E-Commerce - Education - Customer ServiceThe release of InternVL3 is a promising sign for the future of open-source MLLMs. By providing powerful and freely accessible models, the innovative power of the community is fostered and the development of new applications is accelerated. This can help reduce dependence on commercial providers and make the development of AI technologies more transparent and accessible.
The architecture of InternVL3 is based on the latest research results in the field of deep learning and utilizes innovative techniques for processing multimodal data. The developers have placed great emphasis on efficiency and scalability to make the model accessible to users with limited resources. The release of the model under an open-source license allows the community to examine, modify, and further develop the code.
The development of InternVL3 and other open-source MLLMs is an important step towards a future where AI technologies are accessible to everyone. The increasing performance and availability of these models will further drive the development of new applications and the integration of AI into various areas of life. It remains exciting to observe how the landscape of multimodal AI will evolve in the coming years and what contribution open-source projects like InternVL3 will make.
Mindverse, a German provider of AI solutions, offers a comprehensive portfolio of tools and services for the development and implementation of AI applications. From text generation and image editing to the development of customized chatbots, voicebots, AI search engines, and knowledge systems, Mindverse supports companies in optimally exploiting the potential of artificial intelligence. With a focus on innovation and quality, Mindverse accompanies its customers on their journey into the future of AI.
Bibliographie: https://x.com/opengvlab/status/1912109914379718720 https://twitter.com/HuggingPapers/status/1912220230383813061 https://mmmu-benchmark.github.io/ https://huggingface.co/OpenGVLab/InternVL3-8B https://www.aibase.com/news/www.aibase.com/news/17079 https://arxiv.org/html/2404.16821v2 https://www.researchgate.net/publication/387248634_How_far_are_we_to_GPT-4V_Closing_the_gap_to_commercial_multimodal_models_with_open-source_suites https://github.com/BradyFU/Awesome-Multimodal-Large-Language-Models