The Allen Institute for AI has developed a new multimodal AI model called Molmo that has the ability to analyze and describe images, as well as point to different parts of the image. Molmo consists of four models with different sizes, ranging from 1 billion to 72 billion parameters. This open approach to AI demonstrates the power of high-quality training data and unlocks new capabilities for AI agents, robots, and augmented and virtual reality applications. Molmo was able to successfully identify various items in a refrigerator, showcasing its impressive visual perception abilities.
Ai2’s CEO, Ali Farhadi, highlighted the institute’s open approach to artificial intelligence, releasing training data, annotations, code, and other data for researchers and developers to use. Molmo demonstrates that open models can compete with proprietary alternatives on key performance benchmarks. The smaller models within Molmo are performing as well as larger models, enabling them to run on a wider range of devices. The Molmo models compare favorably with those from other companies, such as OpenAI, Anthropic, and Google.
Molmo’s training on high-quality, curated data sets it apart from other models that rely on large, noisy web crawled datasets. By using careful human annotations, Molmo achieves higher accuracy and reliability in its visual recognition capabilities. The model can identify details in complex images, count the number of objects, and perform tasks such as telling time from a traditional clock face. These breakthroughs in visual recognition open up possibilities for developers to create new forms of AI agents that can perform tasks like reading web pages and placing orders online.
The Allen Institute for AI, founded by Paul Allen, has been led by Ali Farhadi for over a year. The institute focuses on developing AI breakthroughs that serve society and offers its technology for public use. Molmo is the latest AI model from Ai2, following the release of the Open Language Model (OLMo) last year, which won the Innovation of the Year award at the 2024 GeekWire Awards. While Ai2 typically does not develop its own products, it is taking a more public approach with the Molmo demo site to bring attention to the technology and further its mission of advancing AI for societal benefit.
Molmo’s abilities were highlighted in demos at the Ai2 headquarters, showcasing its visual recognition capabilities in identifying details in images and reading web pages. The model’s unique features, such as the ability to count dogs with their tongues out and tell time from clock faces, set it apart from other AI models. Molmo also has the potential to be used in creating autonomous AI agents that can perform tasks like browsing websites and placing orders. As an experiment in public demonstration of AI technology, Ai2 aims to gather feedback on Molmo’s capabilities and impact on the AI research community.
