Post-Training at the Frontier: RL Post-Training Nemotron 3.5 Lightning
CoreWeave, Thursday, October 22nd, 2026
CoreWeave details a three-week RL post-training sprint with NVIDIA and You.com that lifted BrowseComp accuracy to 45.45%.
CoreWeave describes a three-week sprint with NVIDIA and You.com to post-train an open-weight Nemotron model for agentic web search using web-grounded data, supervised fine-tuning, reinforcement learning and managed RL Rollouts.
The final workflow reached 45.45% accuracy on BrowseComp versus a 36.97% baseline, a 22.94% relative improvement, while using 31.55% fewer tool calls per task.
The company says inference cadence is the main bottleneck in RL loops and that CoreWeave RL Rollouts, built with NVIDIA Dynamo, reduces the overhead of syncing updated weights to serving replicas.
The post covers You.com's data strategy, difficulty filtering and GRPO training, plus practical lessons.