NVIDIA Fine-Tunes Nemotron for Gold-Level Results at IOI and IMO 2026

NVIDIA demonstrated that the Nemotron model family can be fine-tuned into world-class specialists for both the International Olympiad in Informatics (IOI) and the International Mathematical Olympiad (IMO). By combining supervised fine-tuning (SFT), reinforcement learning (RL), and feedback-driven inference, the teams achieved gold-medal-level performance in both competitions.

IOI 2026 Performance

The Nemotron-3-Ultra-CC model, with 550 billion total parameters and 55 billion active parameters, achieved a score of 535.4/600 in a live, prospective run. This result exceeded the official gold threshold of 361.12 and surpassed the top human score of 498.27. The system utilized SFT and the GenCorrect iterative generate-evaluate-refine strategy. A smaller variant, Nemotron-3-Nano-CC (30B total/3B active), also crossed the gold threshold of 438.3 with a score of 468 after applying SFT, RL, and GenCorrect.

IMO 2026 Performance

For the IMO, NVIDIA trained specialists on Nemotron 3 Ultra using SFT and RL. The SFT corpus included 414,890 quality-filtered examples across 15,818 unique proof problems, while the RL model was trained on 9,597 problems near the capability frontier. The final system, which combined SFT and RL checkpoints in a generate-verify-refine loop without external tools or internet access, scored 30/42 points. This exceeded the official gold-medal threshold of 29, including full credit on four of the six problems.

Open Resources

NVIDIA has released the models, data, and recipes on Hugging Face. Key resources include the Nemotron-3-Ultra-CC model, the Nemotron Labs IMO 2026 collection with training datasets and Nemotron-IMO-Bench (200 problems), and the NeMo-Skills repository containing inference pipelines and prompts.