NVIDIA Releases Nemotron-Cascade 2: Compact 30B MoE Model Achieves Gold Medal Performance in IMO and IOI
Key Takeaways
- ▸Nemotron-Cascade 2 achieves Gold Medal-level IMO and IOI performance with only 3B activated parameters, demonstrating unprecedented intelligence density for open-weight models
- ▸Expanded Cascade RL and novel multi-domain on-policy distillation techniques enable competitive reasoning performance in a compact MoE architecture
- ▸Open-weight release of model checkpoints and training data democratizes access to advanced post-training techniques and high-capability reasoning models
Summary
NVIDIA has introduced Nemotron-Cascade 2, an open-weight 30B mixture-of-experts (MoE) model with just 3B activated parameters that delivers competitive reasoning and agentic capabilities despite its compact size. The model achieves remarkable performance on prestigious benchmarks, becoming only the second open-weight LLM to earn Gold Medal-level scores in the 2025 International Mathematical Olympiad (IMO), International Olympiad in Informatics (IOI), and ICPC World Finals—accomplishing this with 20x fewer parameters than its nearest competitor. This breakthrough demonstrates exceptional "intelligence density" by combining mathematical and coding reasoning capabilities typically found only in much larger frontier models.
The model's success stems from two key technical advancements over its predecessor, Nemotron-Cascade 1. First, NVIDIA significantly expanded Cascade Reinforcement Learning to cover a broader spectrum of reasoning and agentic tasks. Second, the team introduced multi-domain on-policy distillation from the strongest intermediate teacher models for each domain throughout the training process, allowing efficient recovery from benchmark regressions while maintaining strong performance gains. NVIDIA has released both the model checkpoints and training data to the open-source community, advancing research in efficient post-training methodologies.
Editorial Opinion
Nemotron-Cascade 2 represents a significant milestone in making frontier-level reasoning capabilities accessible through more efficient models. The achievement of IMO/IOI Gold Medals with 20x fewer parameters than competitors challenges assumptions about the compute requirements for advanced reasoning and suggests the post-training methodology may be as important as raw model scale. However, the practical applicability of this 30B MoE model for real-world agentic tasks beyond specialized math and coding domains remains to be seen, and the release's true impact will depend on community adoption and reproducibility.
