Nemotron Cascade 2 30B A3B
NVIDIA · 2026-03-19 · 31.6B parameters
Nemotron Cascade 2 30B A3B is NVIDIA's open-weight post-trained MoE model based on Nemotron-3-Nano-30B-A3B-Base; NVIDIA labels it 30B total/3B activated, while the released checkpoint contains exactly 31,577,937,344 parameters and no exact active-parameter count is published. It was trained with SFT, Cascade RL, and multi-domain on-policy distillation for reasoning and agentic tasks, supports switchable thinking and instruct modes, and is documented for context lengths up to 1M tokens although its checkpoint configuration defaults to 262,144. Its 52-layer hybrid architecture combines 23 Mamba-2, 23 MoE, and six GQA layers without positional embeddings, and the weights are released under the NVIDIA Open Model License.