CoreWeave RL Rollouts hot-loads weights 15x faster, lifts BrowseComp to 45.45%
_ScottCondron · x · 2026-10-19
CoreWeave launched RL Rollouts: in RL, inference is part of the training loop, and each checkpoint used to mean a redeploy. RL Rollouts hot-loads new weights into a live deployment without touching in-flight requests, 15x faster than a redeploy cycle. NVIDIA and you.com used it to post-train Nemotron 3.5 Lightning for web search, lifting BrowseComp accuracy from 36.97% to 45.45% while cutting tool calls by 30.24%. It's part of CoreWeave's effort to close the loop between production inference and post-training.
More from Infra
- 4 free CUDA resources to learn GPU programming from scratch, from NVIDIA to Oxford — techNmak · 2026-10-19
- Running Moss image model on a 4GB RTX 3050, optimized down to 62s per image — sanketshinde2004 · 2026-10-19
- Barclays: pure-play data center bond spreads hit ~209bps as market prices in contagion — SumitGup · 2026-10-19
- GPU programmer says the field needs lighter, more predictable inference engines — brandon_xyzw · 2026-10-19
- llama-doctor: CLI that diagnoses your llama.cpp server from /metrics with severity-graded findings — msalsas · 2026-10-19
- Underwater AI data center planned in Alaska's Cook Inlet, home to 330 endangered belugas — anselm · 2026-10-19