We extend autoregressive multi-agent sequence models, including Sable and MAT, to the Offline MARL setting and demonstrate that they significanlty outperform current state-of-the-art methods across a diverse set of benchmarks with up to 50 agents.