Nemotron Achieves Gold-Level Scores in IOI and IMO Competitions

Nemotron-3-Ultra-CC scored 535.4 out of 600 in the 2026 International Olympiad in Informatics (IOI), surpassing the gold threshold of 361.12. This score exceeds the top human score of 498.27, demonstrating the model's competitive edge. In the 2026 International Mathematical Olympiad (IMO), the Nemotron 3 Ultra achieved a score of 30 out of 42, also above the gold threshold of 29.

The Nemotron family includes models with varying parameters: the Nemotron-3-Nano-CC features 30 billion total parameters with 3 billion active, while the Nemotron-3-Ultra-CC boasts 550 billion total parameters and 55 billion active. These models were fine-tuned using the GenCorrect iterative generate-evaluate-refine strategy, which has proven effective in enhancing performance.

In the IOI 2025 competition, the Nemotron-3-Nano-CC's score improved significantly from 130 points before supervised fine-tuning (SFT) to 468 points after applying GenCorrect, surpassing the gold threshold of 438.3. The model's score in IOI 2026 was further refined to 502 points after additional training.

The SFT corpus for the IMO consisted of 414,890 examples across 15,818 unique proof problems, while the reinforcement learning (RL) model was trained on 9,597 proof problems. NVIDIA claims that these results indicate Nemotron's adaptability as a foundation for building specialist models, emphasizing that the system operates solely in natural language without the need for formal provers or external tools.

Hugging Face is the platform where the models are available, providing resources for further development and application. This follows previous coverage on the adoption of open models for AI strategies by telecom operators and the launch of new AI workloads by CoreWeave and NVIDIA.