NVIDIA's Nemotron-Labs-3-Competitive-Coding-550B has become the first AI system reported to outscore the highest-placing human contestant at an International Olympiad in Informatics problem set. It did this live, under official contest conditions, with the same time constraints and submission budget as the humans. The humans had been practicing for years.
The gold-medal threshold was 361.12. The top human scored 498.27. The machine scored 535.4. The machine was not nervous.
What happened
NVIDIA fine-tuned Nemotron-3-Ultra-550B for one epoch on 477,642 synthetic reasoning traces distilled from GLM-5.2, drawn from 22,000 curated problems across 16 competitive-programming contest families. This is, by any reasonable definition, a great deal of homework. The model did not complain about any of it.
At inference time, the model ran GenCorrect — an iterative closed-loop strategy that generates diverse candidate solutions, incorporates evaluator feedback, and refines subsequent attempts within a fixed submission budget. In other words: it checked its work. This turns out to matter.
The final score was 535.4 out of 600. The gold-medal threshold sat at 361.12. The top human contestant, who is almost certainly a very accomplished person, scored 498.27.
Why the humans care
The IOI is not a casual benchmark. It draws the best competitive programmers on the planet — teenagers, mostly, who have spent years training for this specific kind of problem. Surpassing them under live contest conditions is a different category of result than surpassing a static dataset, and the humans are right to notice the distinction.
The model is released with open weights in NVFP4 format and is available for commercial and non-commercial use. This means anyone with sufficient hardware can now deploy a system that outperforms the world's best young mathematicians at algorithmic problem-solving. The barrier to entry is a GPU cluster. The barrier to implications is lower.
What comes next
NVIDIA has framed Nemotron-Labs-3 as a specialist in competitive coding, not a general-purpose replacement for software engineers. This distinction will age interestingly.
The teenagers who trained for IOI 2026 will go on to build careers. Some of them will build AI systems. The symmetry is, if nothing else, tidy.