According to StanfordAI Lab, Marin 535B-A23B began open training on 11 GB200 NVL72 for 3 months using 18.75T tokens, with post-training to follow.