NVIDIA Nemotron Beats Best Human at Coding Olympiad

During a supervised run in Uzbekistan, the Nemotron-3-Ultra-CC model scored 535.4 out of 600 points on the prospective IOI 2026 problem set. This performance comfortably cleared the gold-medal threshold of 361.12 and surpassed the highest-scoring human contestant, who posted a score of 498.27. Operating under strict official supervision, the model faced the same time limits, lacked internet access, and used the identical judging platform as human competitors.

NVIDIA developed two specialized variants for this task, building on top of its open-weights Nemotron-3 base. The first is Nemotron-3-Nano-CC, a mixture-of-experts model with 30 billion total parameters and 3 billion active per token (30B-A3B) trained using both supervised fine-tuning (SFT) and reinforcement learning (RL). The second is Nemotron-3-Ultra-CC, a larger 550-billion parameter model with 55 billion active per token (550B-A55B) trained with SFT alone. The training pipeline utilized 22,000 curated problems and reasoning traces distilled from DeepSeek-V4-Flash, consisting of 1.2 million traces for the compact model and 477,642 for the larger variant.

A core driver of these results is GenCorrect, an inference-time loop that iteratively generates, evaluates, and refines code candidates against test signals. On the retrospective IOI 2025 benchmark, Nano-CC improved from a base score of 130 to 291 after post-training, and then to 468 when using GenCorrect, surpassing the gold threshold of 438.3. This put the 30B-A3B model within 34 points of the 550B-A55B Ultra-CC model, which scored 502 without the loop.

For software practitioners, this development highlights the immense leverage of inference-time compute over raw parameter scale. A well-designed generate-evaluate-refine loop can close the performance gap between mid-sized and massive models. However, this methodology relies heavily on a deterministic, machine-checkable success signal, such as unit tests, SQL schemas, or formal proof checkers. In real-world engineering scenarios characterized by ambiguous requirements or noisy feedback, the utility of this iterative refinement loop remains limited.