Nemotron‑Cascade 2
Nemotron‑Cascade 2 surfaced as a 30B MoE model using Cascade RL plus multi‑domain distillation — the team touts ‘Gold Medal’ math competition results and a high ‘intelligence density.’ (x.com)
NVIDIA posted the project page for Nemotron‑Cascade 2 on March 16, 2026, listing the paper, model weights and a team of 17 contributors with equal‑contribution markers for several authors. (research.nvidia.com/labs/nemotron/nemotron-cascade-2/) The group submitted the arXiv preprint (arXiv:2603.19220) on March 19, 2026; the arXiv record names Zhuolin Yang as first author and lists co‑authors such as Bryan Catanzaro, Mohammad Shoeybi and Wei Ping. (arxiv.org/abs/2603.19220) The team published model checkpoints and training data on Hugging Face and released the project under the NVIDIA Open Model License, with the model card showing community metadata (about 97 likes and NVIDIA’s 52.8k follower count on the page). (huggingface.co/nvidia/Nemotron-Cascade-2-30B-A3B) Benchmark numbers in the model card include an IMO AnswerBench score of 79.3, IMO ProofBench 72.9, an IOI 2025 aggregate score listed as 439.3, and an ICPC World Finals result shown as 10/12 on the table. (huggingface.co/nvidia/Nemotron-Cascade-2-30B-A3B) The release notes say the model is post‑trained from a Nemotron‑Nano‑V3 base, document a staged RL + on‑policy distillation training pipeline, and report that this post‑training recovered prior benchmark regressions while improving reasoning and agentic metrics. (research.nvidia.com/labs/nemotron/nemotron-cascade-2/) Operational details on the model card include a “thinking” mode invoked with <think>…</think> tags, tool‑integration results reported in brackets, and recommended sampling settings of temperature = 1.0 and top_p = 0.95 for interactive use. (huggingface.co/nvidia/Nemotron-Cascade-2-30B-A3B)
