Nvidia's Open-Source Post-Training Recipe Achieves Gold-Medal AI With 3B Parameters
The Development
Nvidia released Nemotron-Cascade 2 on March 23, 2026 — an open-weight 30B Mixture-of-Experts model that activates only 3B parameters at inference time. Despite that compact footprint, it achieved gold-medal-level performance on the 2025 International Mathematical Olympiad, the International Olympiad in Informatics, and the ICPC World Finals — only the second open model to reach that tier, after DeepSeek-V3.2-Speciale, which uses 20 times more active parameters. On LiveCodeBench v6, Nemotron-Cascade 2 scores 87.2, surpassing models up to 1 trillion parameters. The post-training pipeline behind these results — Cascade RL, a sequential domain-by-domain reinforcement learning approach, combined with Multi-Domain On-Policy Distillation (MOPD) — is fully open-source via Nvidia's Nemo-RL repository. Notably, the model starts from the same base as Nemotron-3-Nano yet outperforms Nemotron-3-Super, which carries four times the active parameters. Benchmark scores are self-reported by Nvidia.
Our Take
The real release here isn't the model — it's the methodology. Nvidia has published a reproducible blueprint for building domain-specific reasoning systems without expensive pre-training from scratch. Cascade RL's sequential domain ordering means enterprises can bolt new capabilities onto existing models without rebuilding pipelines, while MOPD eliminates the need for costly external teacher models by distilling from a team's own best intermediate checkpoints. This shifts the competitive calculus for enterprise AI teams: raw model size is no longer the primary lever. A team with a disciplined post-training pipeline can close the performance gap on targeted tasks — financial modeling, compliance analysis, software engineering — at a fraction of frontier infrastructure costs. The honest benchmark disclosures (the model underperforms on knowledge retrieval and agentic tasks) also signal which use cases this approach actually serves, which is more strategically useful than blanket capability claims.
The Exploit
Action Item
If your team is applying reinforcement learning to improve a model across multiple capabilities simultaneously, restructure to Cascade RL's sequential domain ordering before your next training run — mixed-domain RL is actively costing you performance on every front.