With backing from SURF and NWO, Dheeraj Varghese and Mohammad Derakhshani—researchers at the University of Amsterdam—accessed the LUMI supercomputer to build NeoBabel, an image-generation system responding to text prompts in multiple languages. The project targeted six linguistically distinct languages: English, Dutch, Chinese, Hindi, French and Persian.
LUMI's computational capacity proved essential to the effort. The researchers enlarged their training dataset to 124 million image-label pairs—up from an initial 40 million—and achieved training at a scale that would have been unattainable using computing infrastructure available within the Netherlands alone. The resulting model has been released publicly and works as intended.
NeoBabel operates under an open-source license, allowing researchers and developers to access, modify and extend the technology. The team positioned the release as a step toward making generative AI systems more accessible across different languages and cultures.
