Nemotron-Cascade 2: Open 30B MoE Delivers Best-in-Class Reasoning and Agentic Performance at 3B Active Params

arXiv·high signal

NVIDIA released Nemotron-Cascade 2 on arXiv today — a 30B MoE model with only 3B activated parameters, trained with a novel Cascade RL + multi-domain on-policy distillation pipeline. The paper claims best-in-class reasoning scores among open models at its efficiency tier and strong agentic task performance, making it a serious option for cost-sensitive production agent deployments. The Cascade RL approach is the key architectural contribution: it progressively distills larger teacher reasoning chains into the MoE routing policy rather than the dense weights.