SenseTime's NEO Cuts Multimodal Training Data by 90%, Opening Custom AI to Resource-Constrained Enterprises

SenseTime has open-sourced NEO, a multimodal architecture requiring 390 million image-text pairs for training—one-tenth the volume of comparable models at equivalent performance. The 2B and 9B parameter variants enable enterprises to build domain-specific models for medical imaging and industrial automation without acquiring massive datasets. The approach reflects broader industry momentum toward architectural efficiency as an alternative to raw parameter scaling.