Nemotron-Cascade 2: Post-Training LLMs with Cascade RL and Multi-Domain On-Policy Distillation
Zhuolin Yang, Zihan Liu, Yang Chen, Wenliang Dai, Boxin Wang, Sheng-Chieh Lin, Chankyu Lee, Yangyi Chen, Dongfu Jiang, Jiafan He, Renjie Pi, Grace Lam, Nayeon Lee, Alexander Bukharin, Mohammad Shoeybi, Bryan Catanzaro, Wei Ping
2026-03-20
Summary
This paper introduces Nemotron-Cascade 2, a new artificial intelligence model that's surprisingly powerful for its size. It's designed to be good at complex thinking, like math and coding, and also at acting as an 'agent' to complete tasks.
What's the problem?
Existing large language models (LLMs) are often enormous, requiring huge amounts of computing power. While these models perform well, their size makes them difficult for many researchers and developers to use. The challenge is to create a model that can achieve similar levels of intelligence with significantly fewer resources.
What's the solution?
The researchers built Nemotron-Cascade 2 using a technique called 'Mixture of Experts,' which allows it to be efficient. They started with a carefully chosen dataset and then used a process called 'Cascade RL' – essentially training the model through rewards and feedback – across many different areas of reasoning and problem-solving. A key improvement was using stronger 'teacher' models to guide the learning process, helping the model recover from mistakes and consistently improve its performance.
Why it matters?
Nemotron-Cascade 2 is important because it demonstrates that you don't necessarily need a massive model to achieve high-level intelligence. It performs exceptionally well on challenging competitions like the International Mathematical Olympiad and coding contests, rivaling much larger models. This opens the door for more accessible and efficient AI research and development, allowing more people to build and experiment with powerful AI tools.
Abstract
We introduce Nemotron-Cascade 2, an open 30B MoE model with 3B activated parameters that delivers best-in-class reasoning and strong agentic capabilities. Despite its compact size, its mathematical and coding reasoning performance approaches that of frontier open models. It is the second open-weight LLM, after DeepSeekV3.2-Speciale-671B-A37B, to achieve Gold Medal-level performance in the 2025 International Mathematical Olympiad (IMO), the International Olympiad in Informatics (IOI), and the ICPC World Finals, demonstrating remarkably high intelligence density with 20x fewer parameters. In contrast to Nemotron-Cascade 1, the key technical advancements are as follows. After SFT on a meticulously curated dataset, we substantially expand Cascade RL to cover a much broader spectrum of reasoning and agentic domains. Furthermore, we introduce multi-domain on-policy distillation from the strongest intermediate teacher models for each domain throughout the Cascade RL process, allowing us to efficiently recover benchmark regressions and sustain strong performance gains along the way. We release the collection of model checkpoint and training data.