NVIDIA's Nemotron model has achieved a significant milestone by earning gold-medal performance at both the 2026 International Olympiad in Informatics (IOI) and International Mathematical Olympiad (IMO), two of the world's most rigorous tests of algorithmic and mathematical reasoning. The IOI-specialized variant scored 535.4 out of 600 points, exceeding both the official gold threshold and the top human score of 498.27. A separate IMO-focused system scored 30 out of 42 points, surpassing the gold-medal threshold of 29 and achieving full credit on four of six problems, with all submissions graded by official IMO evaluators. The achievement emerged from a methodical specialization approach rather than raw computational scaling. NVIDIA's team started with Nemotron 3 foundation models and applied supervised fine-tuning, reinforcement learning, and domain-specific training data—curating 22,000 competitive programming problems and 414,890 quality-filtered mathematical proof examples. Critically, they paired these specialized models with inference strategies that iteratively generate candidate solutions, evaluate them, and refine the most promising attempts. For IOI, the GenCorrect feedback loop enabled dramatic improvements across multiple refinement rounds. For IMO, complementary SFT and RL checkpoints worked together to both generate and critique mathematical proofs entirely in natural language. The results underscores a core principle: exceptional AI performance emerges from co-designing the model, the data, and the inference loop. Fine-tuning alone did not produce the medals, nor did raw sampling. The reusable recipe—starting with a strong foundation, curating domain-specific training data, applying post-training methods, and pairing with feedback-driven inference—offers a template for adapting general-purpose models to specialized and demanding domains.