Nvidia says its Nemotron-3-Ultra-CC model outscored the top human at the 2026 Olympiad in Informatics
Running live under the same no-internet, local-execution rules as the teenagers competing against it, the 550-billion-parameter system posted 535.4 out of 600, above both the gold-medal cutoff and the top human score. The claim comes from Nvidia's own unreviewed preprint, with almost no independent verification yet.
Nvidia researchers posted a paper on arXiv on September 2, 2026 claiming that a competition-tuned version of their Nemotron 3 Ultra model, called Nemotron-3-Ultra-CC, outscored every human contestant at the 2026 International Olympiad in Informatics (IOI). It ran live under the same constraints as the teenage competitors: no internet access, submissions judged locally, the same time limits. The system scored 535.4 out of 600, clearing both the contest's gold-medal threshold of 361.12 and the top human score of 498.27.
A different model, a different competition
This is a separate system from "Nemotron-Cascade," the Nvidia model line associated with earlier work around the International Mathematical Olympiad. Nemotron-3-Ultra-CC is newer (550 billion total parameters, 55 billion active), built on Nvidia's Nemotron 3 Ultra base model, and IOI is a programming contest. The paper credits most of the gain to GenCorrect, a test-time strategy that generates, checks, and revises candidate solutions using the contest's automated judge as feedback. The authors write that it's "the first AI system to outscore the highest-scoring human contestant on an IOI problem set."
Nvidia is grading its own model
The paper hasn't been peer reviewed. Nvidia's own team is publishing a claim about Nvidia's own model, and at the time of writing the only outside discussion we found was a single automated analysis blog; no mainstream tech outlet had covered it. None of that makes the arithmetic wrong. IOI's scoring is an objective pass/fail judge, so the result is easy for someone else to check. Until someone outside Nvidia does, read "first to beat a human" as the authors' claim about their own system.
Contest results with a built-in judge, like a compiler or an automated grader, are the rare case where a vendor's claim about its own model can be checked directly. Almost none of the AI work an organization pays for looks like that. A quarterly report or a customer email has no compiler, so a coding-contest score tells you little about what your own AI spend is producing.