Event details
How You.com, NVIDIA, and CoreWeave Improved Nemotron 3.5 Lightning with RL Rollouts
Reinforcement learning moves quickly, but training and rollout serving don’t always move at the same cadence. Policy updates may land every few hours. Getting new checkpoints onto inference workers while keeping serving available adds another infrastructure problem to solve.
Join engineers from You.com, NVIDIA, and CoreWeave for a technical walkthrough of how they customize NVIDIA Nenmotron 3.5 Lightning for complex web research using You.com Search, supervised fine-tuning (SFT), reinforcement learning (RL), and RL Rollouts powered by NVIDIA Dynamo on CoreWeave infrastructure. Then see the mechanisms come together in a worked example with Nemotron 3.5 Lightning on NVIDIA HGX B200. NVIDIA Nemotron 3.5 Lightning is a customizable open model for always-on agents to complete specialized tasks faster with leading domain-specific accuracy across local, on-premises, and cloud deployments. The walkthrough will cover before-and-after traces on the same task, the compute cost and timeline, and the reported results: a 22.94% accuracy gain and 30.24% fewer tool calls. Leave with a blueprint for applying the approach to your own model.
In this webinar, we’ll cover:
- How post-training fits into the model-capability stack, and why search infrastructure matters for RL
- How You.com builds training questions from the web graph, how that differs from open datasets, and why decontamination matters
- How filtering training tasks by difficulty can improve accuracy and training efficiency
- How RL Rollouts handles delta weights, checkpoint hot-loading, and on-policy weight transitions to keep inference synchronized with policy updates
- Where per-token training metadata, including logprobs, expert routing, and token IDs, enters the response contract
- A worked Nemotron 3.5 Lightning example on NVIDIA HGX B200, with before-and-after traces on the same task, compute cost, timeline, and results
Register to see how the pieces fit together and take away a blueprint for your own post-training workflow.
