Nvidia open sources its Cascade RL pipeline via NeMo-RL, enabling enterprises to achieve top-tier reasoning performance without building massive models from scratch.
Nvidia has open-sourced the post-training pipeline behind its Nemotron-Cascade 2 model, signalling a shift in AI development where training methodology, rather than model size, defines performance. Released via the NeMo-RL repository, the Cascade RL framework provides a reproducible approach for enterprises to build high-performance reasoning systems without starting from scratch.
Nemotron-Cascade 2 is an open-weight 30B Mixture-of-Experts model that activates just 3B parameters at inference, yet delivers gold medal-level results across the 2025 International Mathematical Olympiad, International Olympiad in Informatics, and ICPC World Finals. Despite its compact footprint, it outperforms both Nemotron-3-Nano and the significantly larger Nemotron-3-Super.
At the core is Cascade RL, a sequential reinforcement learning pipeline that trains models domain by domain—ranging from instruction-following and multi-domain reasoning to code and software engineering. This approach avoids catastrophic forgetting, enables domain-specific optimisation, and improves compute efficiency.
Complementing this is Multi-Domain On-Policy Distillation (MOPD), which reuses intermediate checkpoints as internal teachers, eliminating the need for external models. Nvidia reports faster convergence and higher efficiency, including recovering benchmark performance in significantly fewer optimisation steps.
The model achieves strong scores across reasoning benchmarks, including 87.2 on LiveCodeBench and 98.6 on AIME 2025 with tool integration, though it trails in knowledge-heavy and agentic tasks.
For enterprises, the implications are significant: modular capability upgrades, lower infrastructure costs, and a practical path to deploy advanced reasoning systems. The release underscores a broader industry shift towards “intelligence density,” where better training pipelines—not larger models—drive AI progress.
