“The FoMo-in-Flux benchmark for continual multimodal pretraining was constructed using 63 datasets and is designed with realistic compute constraints and practical deployment requirements.”