Cambrian-S
modelA spatially-grounded video MLLM family (0.5B–7B) from Saining Xie's group with Stanford, adding predictive-sensing memory for long video understanding. Released with the VSI-590K training data and the VSI-SUPER stress test; weights and data are on HuggingFace (Cambrian-S-3M, ~13K downloads). Yann LeCun is a co-author at his NYU exit. Successor to Cambrian-1 on the VSI-Bench spatial-intelligence agenda.