NVIDIA Releases Nemotron-Cascade 2: An Open 30B MoE with 3B Active Parameters, Delivering Better Reasoning and Strong Agentic Capabilities

2026-03-20 · Source: Machine Learning ML & Generative AI News · Field: Technology & Digital — Artificial Intelligence & Machine Learning · Depth: Advanced, quick

Summary

NVIDIA has released Nemotron-Cascade 2, an open-weight Mixture-of-Experts (MoE) model featuring a 30B architecture with 3B active parameters, designed to enhance "intelligence density." This model is the second open-weight offering to achieve Gold Medal-level performance in IMO 2025 and IOI 2025 benchmarks. Its core innovation lies in integrating Cascade RL with Multi-domain On-Policy Distillation (MOPD), which provides a dense token-level advantage and improves sample efficiency compared to sequence-level reward methods like GRPO. Nemotron-Cascade 2 demonstrates strong performance in math, coding, and instruction following, outperforming Qwen3.5-35B-A3B on AIME 2025 and ArenaHard v2, though it trades off performance in knowledge-intensive tasks. It also features a 1M context window and a "Thinking Mode" for complex reasoning and agentic workflows.

Key takeaway

For AI Architects and Research Scientists evaluating open-weight models for complex reasoning, Nemotron-Cascade 2 offers a compelling option due to its strong performance in math, coding, and agentic capabilities. Consider its 1M context window and "Thinking Mode" for applications requiring deep logical processing, but be mindful of its reduced efficacy in knowledge-intensive domains.

Key insights

NVIDIA's Nemotron-Cascade 2 MoE model excels in reasoning and agentic tasks via Cascade RL and MOPD.

Principles

• MoE architectures can achieve high intelligence density.
• Token-level advantages improve sample efficiency in RL.
• Specialized models may trade off knowledge for reasoning.

Method

Nemotron-Cascade 2 integrates Cascade RL with Multi-domain On-Policy Distillation (MOPD) to provide dense token-level advantages, enhancing sample efficiency and recovering performance regressions during training.

In practice

• Utilize for complex math and coding tasks.
• Employ for instruction following applications.
• Leverage "Thinking Mode" for agentic workflows.

Topics

• Nemotron-Cascade 2
• Mixture-of-Experts
• Reinforcement Learning
• Agentic AI
• Complex Reasoning

Best for: AI Scientist, Research Scientist, AI Architect, AI Researcher, AI Engineer, Machine Learning Engineer

Related on AIssential

• NVIDIA Releases Nemotron 3 Super: A 120B Parameter Open-Source Hybrid Mamba-Attention MoE Model Delivering 5x Higher Throughput for Agentic AI — Nemotron 3 Super is an open-source 120B parameter hybrid MoE model for efficient multi-agent reasoning.
• Nemotron-Cascade 2: Post-Training LLMs with Cascade RL and Multi-Domain On-Policy Distillation — Nemotron-Cascade 2 achieves frontier-level reasoning with 20x fewer parameters via expanded Cascade RL and multi-domain distillation.
• NVIDIA Nemotron 3 Ultra Powers Faster, More Efficient Reasoning for Long-Running Agents — Nemotron 3 Ultra optimizes long-running AI agents with a specialized Mixture-of-Experts model for faster, more cost-effective reasoning.
• MiniMax M2.7 Advances Scalable Agentic Workflows on NVIDIA Platforms for Complex AI Applications — MiniMax M2.7, a 230B-parameter MoE model, offers high efficiency and performance for agentic tasks via NVIDIA optimizations.
• Run NVIDIA Nemotron 3 Nano as a fully managed serverless model on Amazon Bedrock — Nemotron 3 Nano offers an open, efficient, and accurate SLM for specialized agentic AI systems on Amazon Bedrock.
• Together AI at NVIDIA GTC 2026: Explore our latest innovations across research and products — Together AI and NVIDIA are advancing open, agentic, and production-ready AI systems through new model integrations and inference optimizations.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine Learning ML & Generative AI News.