SenseTime's NEO: Training Multimodal AI with 90% Less Data

SenseTime has open-sourced NEO, a multimodal architecture that achieves comparable performance to rivals using just 390 million image-text pairs—a tenth of what competitors require. The 2B and 9B parameter variants let enterprises build custom models for medical imaging and industrial automation without massive datasets. The approach signals industry momentum away from pure scale toward architectural efficiency.