NVIDIA has published Nemotron-Labs-3-Competitive-Coding, a model tuned for competitive programming, on Hugging Face. The model page reports that it scored 535.4 out of 600 on the International Olympiad in Informatics (IOI) 2026 problem set. That was above the top human contestant’s 498.27 and the listed gold-medal threshold of 361.12.
The result was not from a single answer generated by the model alone. NVIDIA says it paired the model with GenCorrect, a method that produces several possible solutions, uses feedback from an evaluator, and then tries to improve its answers within a set submission budget. The company says the evaluation was run prospectively under the contest’s official time, internet-access, and submission rules.
How it was trained
The model is fine-tuned from NVIDIA’s Nemotron-3-Ultra-550B-A55B. NVIDIA says it trained for one pass on 477,642 synthetic reasoning traces distilled from GLM-5.2. Those traces covered 22,000 curated programming problems from 16 regional and international contest families.
NVIDIA says it chose GLM-5.2 as the source for the training traces because it was more accurate and produced generations roughly 30% shorter than a variant trained using DeepSeek-V4-Flash. The model page describes Nemotron as a family of models with open weights, training data, and recipes, and says this coding model is available for commercial and non-commercial use.
What the score shows
The result is notable, but it is a company-reported score for a particular setup: the fine-tuned model plus GenCorrect, tested on one contest problem set. It does not by itself show how the model performs on everyday software work, or how it compares across other coding tests. The page calls it the first AI system reported to outscore the highest-scoring human contestant on an IOI problem set.
Why it matters
For people who use coding AI, the result is evidence that a model paired with repeated testing and revision can perform strongly on difficult programming puzzles. It is not a guarantee that the system will be equally useful for ordinary coding tasks, and the reported result comes from NVIDIA’s model page.