Nvidia says models from its Nemotron 3 family reached gold-medal standard at this year’s International Olympiad in Informatics and International Mathematical Olympiad. The conditions behind the two differ, and the difference is the story.
The maths result is the better attested. Nvidia reports 30 out of 42 at IMO 2026, above the official gold threshold of 29, with full credit on four of six problems, and says official IMO graders marked the proofs. The system worked in natural language, generating, verifying and refining, with no formal prover, external tools or internet access.
The programming result is self-run. A model called Nemotron-3-Ultra-CC scored 535.4 out of 600 at IOI 2026, which Nvidia places above both the 361.12 gold threshold and the top human score of 498.27. Nvidia itself calls this an unofficial, unsupervised benchmark forming no part of the official IOI ranking, although it ran live under the same time, internet and submission constraints as contestants.
The artefacts are published, which makes the claims arguable rather than merely announced: checkpoints, both training sets (414,890 supervised examples across 15,818 proof problems) and a Nemotron-IMO-Bench benchmark sit on Hugging Face, with inference pipelines and the submitted proofs on GitHub. No licence is stated. One oddity: the model whose name says 550B is listed by the Hugging Face widget as 335B.
Source: Nvidia, One Model Family, Two Gold-Level Results: Fine-Tuning Nemotron for IOI and IMO.