NVIDIA Reports Nemotron Results Above the Gold-Medal Threshold at IMO and IOI 2026

NVIDIA published the Nemotron systems’ results on October 7, 2026, in a joint material on the Hugging Face platform. The company describes two Nemotron family projects designed for olympiad mathematics and competitive programming. For the 2026 International Mathematical Olympiad (IMO), it reports a score of 30 out of 42 points; at the 2026 International Olympiad in Informatics (IOI), it reports 535.4 out of 600 points.

Both figures are above the official gold-medal thresholds for those editions. At IMO 2026, the gold threshold was 29 out of 42 points; at IOI 2026, it was 361.12 out of 600 points. NVIDIA also released some artifacts, including the competitive-programming model, data, code, and methodology.

Nemotron system results are not official medals

The results concern solving olympiad problems, not systems competing in the official standings. For IOI, NVIDIA explicitly states that its run was not an official competitor in the ranking. It describes it as an unofficial, unsupervised evaluation under the same time, internet, and submission constraints.

The company claims that the 535.4 out of 600 points surpassed the best human result at IOI 2026. However, this comparison comes from NVIDIA materials; independent confirmation of the specific AI run from IOI organizers is not available.

Similarly, for the mathematics component, NVIDIA and its preprint state that the proofs were evaluated by official IMO graders. However, this information has not been independently confirmed by a statement from the olympiad organizers.

Three checkpoints for mathematics, an open model for code

The preprint on the mathematics system describes a combination of three checkpoints. Their roles were proof generation, verification, and subsequent revision. According to the authors, the system did not use the internet, external tools, or a formal prover when solving IMO problems.

For competitive programming, NVIDIA released the NVIDIA-Nemotron-Labs-3-Competitive-Coding-550B-A55B-NVFP4 checkpoint. The model’s Hugging Face page lists it under the open OpenMDW-1.1 license. The release therefore concerns not only the reported scores, but also materials that make it possible to examine the approach used.

NVIDIA presents the results as combining two specialties within one Nemotron family: mathematical proving and programming under competition constraints. The materials point to the importance of post-training and demanding inference loops for narrowly defined tasks.

What the results do not yet confirm

Technical preprints are not independently peer-reviewed confirmation of reproducibility. By themselves, they also do not prove that the same capabilities transfer beyond the precisely described olympiad procedures and problems.

Another open question is the computational cost of the inference loops. The published scores should therefore be considered separately from the question of what resources were required to achieve them. Independent replication using the available code, data, and checkpoints will also be important, as will any statements from IMO and IOI organizers concerning the evaluation of AI systems and the comparison conditions.

Verified and updated: 10/07/2026 15:21