Nvidia says two systems adapted from its Nemotron 3 model reached gold-medal-level results on the 2026 International Mathematical Olympiad and International Olympiad in Informatics. The projects used supervised fine-tuning, reinforcement learning and inference loops that generated, checked and improved candidate solutions.
The competitions test different abilities. IMO requires rigorous written mathematical proofs, while IOI requires code that passes hidden tests under time and submission limits. Official IMO graders evaluated the model’s submitted proofs. The IOI result came from a live prospective run under the same time, internet and submission constraints as human competitors, but it was unofficial, unsupervised and did not appear in the official ranking.
Nvidia presents the work as evidence that one foundation model can be specialized for distinct technical fields using curated problems and high-quality reasoning traces. The results do not mean the base model automatically performs at that level, and competition tasks are not a complete measure of software engineering or mathematical research. They do offer a more demanding test than static question sets, particularly where a system must produce code or proofs that external evaluators can check.