Decoupled DiLoCo is a distributed pre-training framework that breaks synchronous lock-step barriers by partitioning compute across independent learners that asynchronously communicate with a central synchronizer. The method achieves improved training efficiency in failure-prone environments with zero global downtime while maintaining competitive performance on text and vision tasks.