NVIDIA Releases Nemotron-Cascade 2: 30B Open Model Achieves IMO Gold Medal with Remarkable Parameter Efficiency
Key Takeaways
- ▸Nemotron-Cascade 2 achieves IMO 2025 and IOI 2025 gold-medal performance with only 3B activated parameters—demonstrating 20× better efficiency than comparable frontier models
- ▸NVIDIA's Cascade RL framework, enhanced with multi-domain on-policy distillation, enables effective post-training across complex reasoning and agentic domains with minimal performance regression
- ▸Full model weights, SFT data, and RL training data are open-sourced, democratizing access to state-of-the-art reasoning capabilities
Summary
NVIDIA has released Nemotron-Cascade 2, a 30-billion parameter mixture-of-experts model with only 3 billion activated parameters that achieves gold-medal performance on the 2025 International Mathematical Olympiad (IMO) and International Olympiad in Informatics (IOI) competitions. The model is only the second open-weight LLM to reach this level of performance, demonstrating that advanced reasoning capabilities can be achieved with 20× fewer parameters than expected. This breakthrough challenges the conventional wisdom that frontier-level reasoning requires massive model sizes.
The technical advancement builds on NVIDIA's Cascade RL framework, which has been substantially expanded to cover broader reasoning and agentic domains. A key innovation is multi-domain on-policy distillation, which uses intermediate teacher models during training to recover performance regressions while maintaining strong gains across diverse reasoning tasks. NVIDIA has released the complete model weights, supervised fine-tuning (SFT) dataset, and reinforcement learning training data, enabling community researchers to build upon this work. Across mathematics, code reasoning, and instruction-following benchmarks, Nemotron-Cascade-2 outperforms both the larger Qwen3.5-35B and Nemotron-3-Super-120B models.
- The model outperforms significantly larger competitors on key benchmarks, proving that intelligence density depends more on post-training methodology than parameter count alone
Editorial Opinion
Nemotron-Cascade 2 is a watershed moment for open AI research, proving that gold-medal mathematical reasoning doesn't require 100+ billion parameters. By focusing on smarter training methodology rather than scale, NVIDIA has challenged the cost barriers that have gatekept frontier-level reasoning models. The decision to open-source the full training data—not just weights—is commendable and raises the bar for reproducibility in the field.
