CoreWeave RL Rollouts Cuts Weight Reload to About 15x Faster
Hot-swapping weights into a running deployment removes the redeploy bottleneck that dominates wall-clock time in RL post-training, so teams can iterate on checkpoints far more cheaply. Engineers running RL pipelines or serving infrastructure should evaluate it against their current checkpoint-deploy scripts. The Nemotron 3.5 Lightning result doubles as a concrete case of using it to improve a search agent's benchmark score.
CoreWeave launched RL Rollouts, a feature that loads new model weights into a live inference deployment without redeploying or interrupting in-flight requests. The company says it is about 15x faster than a full redeploy cycle in reinforcement learning training loops. NVIDIA and You.com used it to post-train Nemotron 3.5 Lightning for web search. The model's BrowseComp accuracy rose from 36.97% to 45.45% while tool calls dropped by 30.24%.
In reinforcement learning, inference is part of the training loop. Every checkpoint used to mean a redeploy, and the trainer waited.
CoreWeave RL Rollouts load the new weights into a live deployment without touching in-flight requests, about 15x faster than a redeploy cycle.
@nvidia and @youdotcom used it to post-train Nemotron 3.5 Lightning for web search, lifting BrowseComp accuracy from 36.97% to 45.45% while reducing tool calls by 30.24%.
Full breakdown here: https://t.co/cTcqSRGeCB
