CoreWeave Launches RL Rollouts to Eliminate GPU Idle Time During Model Training Cycles

CoreWeave unveiled RL Rollouts, a new inference capability that hot-loads updated model weights into live deployments, keeping GPUs productive during reinforcement learning training loops. The feature, released September 30 as part of the CoreWeave Forge platform announcement, addresses a persistent inefficiency: when RL training generates new model weights, inference workers must reload the updated checkpoints, leaving GPUs idle and stalling the next round of rollout generation.