NVIDIA has launched Nemotron-Cascade 2, a cutting-edge 30B Mixture-of-Experts (MoE) model featuring 3B active parameters. This model is engineered to enhance 'intelligence density,' enabling superior reasoning abilities while utilizing fewer parameters than leading models. Notably, Nemotron-Cascade 2 is the second open-weight large language model (LLM) to achieve Gold Medal-level results in prestigious competitions such as the 2025 International Mathematical Olympiad and the International Olympiad in Informatics.
The model excels in specific areas, particularly in mathematical reasoning and coding, outperforming competitors like Qwen3.5-35B-A3B and the larger Nemotron-3-Super-120B-A12B. In mathematical reasoning, it achieved scores of 92.4 and 94.6 on key tests, surpassing its rivals. Its coding performance also stands out, leading in benchmarks like LiveCodeBench and IOI 2025.
The model's architecture incorporates advanced techniques such as Cascade Reinforcement Learning and Multi-domain On-Policy Distillation, which enhance its training efficiency and performance. With two operational modes, it can engage in deep reasoning or provide quick responses, making it versatile for various tasks. Overall, Nemotron-Cascade 2 illustrates that high-level reasoning can be achieved at a reduced scale through targeted learning strategies.